The Ultimate WorkstationFor Open Source AI

Devstral is a specialized 23.6 billion parameter language model developed by Mistral AI and All Hands AI, finetuned from Mistral-Small-3.1 for software engineering tasks. The text-only model features a 128,000-token context window and achieves 46.8% on SWE-Bench Verified benchmarks. Released under Apache 2.0 License, it functions as an agentic coding assistant for codebase exploration, multi-file editing, and automated software engineering workflows.

Wan 2.1 T2V 1.3B is an open-source text-to-video generation model developed by Wan-AI, featuring 1.3 billion parameters and utilizing a Flow Matching framework with diffusion transformers. The model supports multilingual text-to-video synthesis in English and Chinese, operates efficiently on consumer GPUs requiring 8.19 GB VRAM, and generates 480P videos with capabilities for image-to-video conversion and text rendering within videos.

Wan 2.1 I2V 14B 720P is a 14-billion parameter image-to-video generation model developed by Wan-AI that converts single images into 720P videos. Built on a unified transformer-based diffusion architecture with a novel 3D causal VAE (Wan-VAE) for spatiotemporal compression, the model supports multilingual text prompts and demonstrates competitive performance in video generation benchmarks while maintaining computational efficiency across various GPU configurations.

Qwen2.5-VL-3B-Instruct is a multimodal large language model developed by Alibaba Cloud featuring 3 billion parameters. The model combines a Vision Transformer encoder with a Qwen2.5-series decoder to process images, videos, and text through dynamic resolution handling and temporal processing capabilities. It supports object detection, OCR, document analysis, video understanding, and computer interface automation, trained on approximately 1.4 trillion tokens across multiple modalities and released under Apache-2.0 license.

Qwen2.5-VL 72B is a 72-billion parameter multimodal generative AI model developed by Alibaba Cloud that integrates vision and language understanding. The model features dynamic resolution processing, temporal video alignment, and architectural enhancements over previous Qwen2-VL versions. It performs object detection, document parsing, video comprehension, OCR across multiple languages, and functions as a visual agent for interactive tasks, trained on over 1.4 trillion tokens.

DeepSeek R1 is a large language model developed by DeepSeek AI that employs a Mixture-of-Experts architecture with 671 billion total parameters and 37 billion activated during inference. The model utilizes reinforcement learning and supervised fine-tuning to enhance reasoning capabilities across mathematics, coding, and logic tasks, achieving competitive performance on benchmarks including 90.8 on MMLU and 97.3 on MATH-500.

DeepSeek V3 is a Mixture-of-Experts language model with 671 billion total parameters and 37 billion activated per token, developed by DeepSeek-AI. The model features Multi-head Latent Attention, auxiliary-loss-free load balancing, and FP8 mixed-precision training. Trained on 14.8 trillion tokens with a 128,000-token context window, it demonstrates competitive performance across reasoning, coding, and mathematical benchmarks while supporting multilingual capabilities and long-context processing.

DeepSeek VL2 Small is a 2.8 billion parameter multimodal vision-language model that uses a Mixture-of-Experts architecture with dynamic tiling for processing high-resolution images. Built on the DeepSeekMoE-16B framework with SigLIP vision encoding, it handles tasks including visual question answering, OCR, document analysis, and visual grounding across multiple languages, achieving competitive performance on benchmarks like DocVQA while maintaining computational efficiency through sparse expert routing.

Phi-4 is a 14-billion parameter decoder-only Transformer language model developed by Microsoft Research that focuses on mathematical reasoning and code generation through curated synthetic data training. The model supports a 16,000-token context window and achieves competitive performance on benchmarks like MMLU (84.8) and HumanEval (82.6) despite its relatively compact size, utilizing supervised fine-tuning and direct preference optimization for alignment.

HunyuanVideo is an open-source video generation model developed by Tencent that supports text-to-video, image-to-video, and controllable video synthesis. The model employs a Transformer-based architecture with a 3D Variational Autoencoder and utilizes flow matching for generating videos at variable resolutions and durations. It features 13 billion parameters and includes capabilities for avatar animation, audio synchronization, and multi-aspect ratio output generation.

CogVideoX 1.5 5B I2V is an image-to-video generation model developed by THUDM using a diffusion transformer architecture with 3D causal variational autoencoder. The model generates temporally coherent videos from input images and text prompts, supporting resolutions up to 1360 pixels and video lengths of 5-10 seconds at 16 fps, trained on 35 million curated video clips.

Qwen2.5-Coder-32B is a 32.5-billion parameter transformer-based language model developed by Alibaba Cloud, specifically designed for programming and code intelligence tasks. The model supports over 92 programming languages and features capabilities in code generation, completion, repair, and reasoning with a 128,000-token context window. Trained on approximately 5.5 trillion tokens of code and instructional data, it demonstrates performance across various coding benchmarks including HumanEval, MBPP, and multilingual programming evaluations.

Llama 3.2 3B is a multilingual instruction-tuned language model developed by Meta with 3 billion parameters and a 128,000-token context window. The model utilizes knowledge distillation from larger Llama variants, Grouped-Query Attention for efficient inference, and advanced quantization techniques optimized for PyTorch's ExecuTorch framework. Supporting eight languages, it targets assistant and agentic applications while enabling deployment in resource-constrained environments.

Qwen 2.5 Math 7B is a 7.62-billion parameter language model developed by Alibaba Cloud that specializes in mathematical reasoning tasks in English and Chinese. The model employs chain-of-thought reasoning and tool-integrated approaches using Python interpreters for computational tasks. It demonstrates improved performance over its predecessor on mathematical benchmarks including MATH, GSM8K, and Chinese mathematics evaluations, achieving 83.6 on MATH using chain-of-thought methods.

Qwen2.5-Coder-7B is a 7.61 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, specialized for code generation and reasoning across 92 programming languages. The model features a 128,000-token context window, supports fill-in-the-middle code completion, and was trained on 5.5 trillion tokens of code and text data, demonstrating competitive performance on coding benchmarks like HumanEval and mathematical reasoning tasks.

Qwen 2.5 14B is a 14.7 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, featuring a 128,000 token context window and support for over 29 languages. The model utilizes advanced architectural components including Grouped Query Attention, RoPE embeddings, and SwiGLU activation, and was pretrained on up to 18 trillion tokens of diverse multilingual data for applications in reasoning, coding, and mathematical tasks.

Qwen 2.5 72B is a 72.7 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, released in September 2024. The model features a 128,000-token context window, supports over 29 languages, and demonstrates strong performance on coding, mathematical reasoning, and knowledge benchmarks. Built with architectural improvements including RoPE and SwiGLU activation functions, it excels at structured data handling and serves as a foundation model for fine-tuning applications.

Command R (08-2024) is a 32-billion parameter generative language model developed by Cohere, featuring a 128,000-token context window and support for 23 languages. The model incorporates Grouped Query Attention for enhanced inference efficiency and specializes in retrieval-augmented generation with citation capabilities, tool use, and multilingual comprehension. It demonstrates improved throughput and reduced latency compared to previous versions while offering configurable safety modes for enterprise applications.

Phi-3.5 Mini Instruct is a 3.8 billion parameter decoder-only Transformer model developed by Microsoft that supports multilingual text generation with a 128,000-token context window. The model demonstrates competitive performance across 22 languages and excels in reasoning, code generation, and long-context tasks, achieving an average benchmark score of 61.4 while maintaining efficient resource utilization.

AuraFlow v0.3 is a 6.8 billion parameter, flow-based text-to-image generative model developed by fal.ai. Built on an optimized DiT architecture with Maximal Update Parametrization, it features enhanced prompt following capabilities through comprehensive recaptioning and prompt enhancement pipelines. The model supports multiple aspect ratios and achieved a GenEval score of 0.703, demonstrating effective text-to-image synthesis across diverse artistic styles and photorealistic outputs.

Stable Fast 3D is a transformer-based generative AI model developed by Stability AI that reconstructs textured 3D mesh assets from single input images in approximately 0.5 seconds. The model predicts comprehensive material properties including albedo, roughness, and metallicity, producing UV-unwrapped meshes suitable for integration into rendering pipelines and interactive applications across gaming, virtual reality, and design workflows.

FLUX.1 [schnell] is a 12-billion parameter text-to-image generation model developed by Black Forest Labs using hybrid diffusion transformer architecture with rectified flow and latent adversarial diffusion distillation. The model generates images from text descriptions in 1-4 diffusion steps, supporting variable resolutions and aspect ratios. Released under Apache 2.0 license, it employs flow matching techniques and parallel attention layers for efficient synthesis.

Mistral Large 2 is a dense transformer-based language model developed by Mistral AI with 123 billion parameters and a 128,000-token context window. The model demonstrates strong performance across multilingual tasks, code generation in 80+ programming languages, mathematical reasoning, and function calling capabilities. It achieves 84% on MMLU, 92% on HumanEval, and 93% on GSM8K benchmarks while maintaining concise output generation.

Mistral NeMo 12B is a transformer-based language model developed collaboratively by Mistral AI and NVIDIA, featuring 12 billion parameters and a 128,000-token context window. The model incorporates grouped query attention, quantization-aware training for FP8 inference, and utilizes the custom Tekken tokenizer for improved multilingual and code compression efficiency. Available in both base and instruction-tuned variants, it demonstrates competitive performance on standard benchmarks while supporting function calling and multilingual capabilities across numerous languages including English, Chinese, Arabic, and various European languages.

Llama 3.1 70B is a transformer-based decoder language model developed by Meta with 70 billion parameters, trained on approximately 15 trillion tokens with a 128K context window. The model supports eight languages and demonstrates competitive performance across benchmarks for reasoning, coding, mathematics, and multilingual tasks. It is available under the Llama 3.1 Community License Agreement for research and commercial applications.

Gemma 2 9B is a decoder-only transformer language model developed by Google with 9.24 billion parameters, trained on 8 trillion tokens of diverse text data including web documents, mathematical content, and code. The model utilizes bfloat16 precision and supports efficient inference through quantization and acceleration techniques, achieving 71.3% on MMLU and 40.2% on HumanEval benchmarks while being suitable for resource-constrained deployment environments.

DeepSeek Coder V2 Lite is an open-source Mixture-of-Experts code language model featuring 16 billion total parameters with 2.4 billion active parameters during inference. The model supports 338 programming languages, processes up to 128,000 tokens of context, and demonstrates competitive performance on code generation benchmarks including 81.1% accuracy on Python HumanEval tasks.

Qwen2-72B is a 72.71 billion parameter Transformer-based language model developed by Alibaba Cloud, featuring Group Query Attention and SwiGLU activation functions. The model demonstrates strong performance across diverse benchmarks including MMLU (84.2), HumanEval (64.6), and GSM8K (89.5), with multilingual capabilities spanning 27 languages and extended context handling up to 128,000 tokens for specialized applications.

Yi 1.5 34B is a 34.4 billion parameter decoder-only Transformer language model developed by 01.AI, featuring Grouped-Query Attention and SwiGLU activations. Trained on 3.1 trillion bilingual tokens, it demonstrates capabilities in reasoning, mathematics, and code generation, with variants supporting up to 200,000 token contexts and multimodal understanding through vision-language extensions.

DeepSeek V2 is a large-scale Mixture-of-Experts language model with 236 billion total parameters, activating only 21 billion per token. It features Multi-head Latent Attention for reduced memory usage and supports context lengths up to 128,000 tokens. Trained on 8.1 trillion tokens with emphasis on English and Chinese data, it demonstrates competitive performance across language understanding, code generation, and mathematical reasoning tasks while achieving significant efficiency improvements over dense models.

Phi-3 Mini Instruct is a 3.8 billion parameter instruction-tuned language model developed by Microsoft using a dense decoder-only Transformer architecture. The model supports a 128,000 token context window and was trained on 4.9 trillion tokens of high-quality data, followed by supervised fine-tuning and direct preference optimization. It demonstrates competitive performance in reasoning, mathematics, and code generation tasks among models under 13 billion parameters, with particular strengths in long-context understanding and structured output generation.

Llama 3 8B is an open-weights transformer-based language model developed by Meta, featuring 8 billion parameters and trained on over 15 trillion tokens. The model utilizes grouped-query attention and a 128,000-token vocabulary, supporting 8,192-token context lengths. Available in both pretrained and instruction-tuned variants, it demonstrates capabilities in text generation, code completion, and conversational tasks across multiple languages.

Command R+ v01 is a 104-billion parameter open-weights language model developed by Cohere, optimized for retrieval-augmented generation, tool use, and multilingual tasks. The model features a 128,000-token context window and specializes in generating outputs with inline citations from retrieved documents. It supports automated tool calling, demonstrates competitive performance across standard benchmarks, and includes efficient tokenization for non-English languages, making it suitable for enterprise applications requiring factual accuracy and transparency.

Command R v01 is a 35-billion-parameter transformer-based language model developed by Cohere, featuring retrieval-augmented generation with explicit citations, tool use capabilities, and multilingual support across ten languages. The model supports a 128,000-token context window and demonstrates performance in enterprise applications, multi-step reasoning tasks, and long-context evaluations, though it requires commercial licensing for enterprise use.

Playground v2.5 Aesthetic is a diffusion-based text-to-image model that generates images at 1024x1024 resolution across multiple aspect ratios. Developed by Playground and released in February 2024, it employs the EDM training framework and human preference alignment techniques to improve color vibrancy, contrast, and human feature rendering compared to its predecessor and other open-source models like Stable Diffusion XL.

Stable Cascade Stage B is an intermediate latent super-resolution component within Stability AI's three-stage text-to-image generation system built on the Würstchen architecture. It operates as a diffusion model that upscales compressed 16×24×24 latents from Stage C to 4×256×256 representations, preserving semantic content while restoring fine details. Available in 700M and 1.5B parameter versions, Stage B enables the system's efficient 42:1 compression ratio and supports extensions like ControlNet and LoRA for enhanced creative workflows.

Stable Video Diffusion XT 1.1 is a latent diffusion model developed by Stability AI that generates 25-frame video sequences at 1024x576 resolution from single input images. The model employs a three-stage training process including image pretraining, video training on curated datasets, and high-resolution finetuning, enabling motion synthesis with configurable camera controls and temporal consistency for image-to-video transformation applications.

Qwen 1.5 72B is a 72-billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team and released in February 2024. The model supports a 32,768-token context window and demonstrates strong multilingual capabilities across 12 languages, achieving competitive performance on benchmarks including MMLU (77.5), C-Eval (84.1), and GSM8K (79.5). It features alignment optimization through Direct Policy Optimization and Proximal Policy Optimization techniques, enabling effective instruction-following and integration with external systems for applications including retrieval-augmented generation and code interpretation.

The SDXL Motion Model is an AnimateDiff-based video generation framework that adds temporal animation capabilities to existing text-to-image diffusion models. Built for compatibility with SDXL at 1024×1024 resolution, it employs a plug-and-play motion module trained on video datasets to generate coherent animated sequences while preserving the visual style of the underlying image model.

Phi-2 is a 2.7 billion parameter Transformer-based language model developed by Microsoft Research and released in December 2023. The model was trained on approximately 1.4 trillion tokens using a "textbook-quality" data approach, incorporating synthetic data from GPT-3.5 and filtered web sources. Phi-2 demonstrates competitive performance in reasoning, language understanding, and code generation tasks compared to larger models in its parameter class.

Mixtral 8x7B is a sparse Mixture of Experts language model developed by Mistral AI and released under the Apache 2.0 license in December 2023. The model uses a decoder-only transformer architecture with eight expert networks per layer, activating only two experts per token, resulting in 12.9 billion active parameters from a total 46.7 billion. It demonstrates competitive performance on benchmarks including MMLU, achieving multilingual capabilities across English, French, German, Spanish, and Italian while maintaining efficient inference speeds.

Playground v2 Aesthetic is a latent diffusion text-to-image model developed by playgroundai that generates 1024x1024 pixel images using dual pre-trained text encoders (OpenCLIP-ViT/G and CLIP-ViT/L). The model achieved a 7.07 FID score on the MJHQ-30K benchmark and demonstrated a 2.5x preference rate over Stable Diffusion XL in user studies, focusing on high-aesthetic image synthesis with strong prompt alignment.

Stable Video Diffusion XT is a generative AI model developed by Stability AI that extends the Stable Diffusion architecture for video synthesis. The model supports image-to-video and text-to-video generation, producing up to 25 frames at resolutions supporting 3-30 fps. Built on a latent video diffusion architecture with over 1.5 billion parameters, SVD-XT incorporates temporal modeling layers and was trained using a three-stage methodology on curated video datasets.

Yi 1 34B is a bilingual transformer-based language model developed by 01.AI, trained on 3 trillion tokens with support for both English and Chinese. The model features a 4,096-token context window and demonstrates competitive performance on multilingual benchmarks including MMLU, CMMLU, and C-Eval, with variants available including extended 200K context and chat-optimized versions released under Apache 2.0 license.

MusicGen is a text-to-music generation model developed by Meta's FAIR team as part of the AudioCraft library. The model uses a two-stage architecture combining EnCodec neural audio compression with a transformer-based autoregressive language model to generate musical audio from textual descriptions or melody inputs. Trained on approximately 20,000 hours of licensed music, MusicGen supports both monophonic and stereophonic outputs and demonstrates competitive performance in objective and subjective evaluations against contemporary music generation models.

Vocos is a neural vocoder developed by GemeloAI that employs a Fourier-based architecture to generate Short-Time Fourier Transform spectral coefficients rather than directly modeling time-domain waveforms. The model supports both mel-spectrogram and neural audio codec token inputs, operates under the MIT license, and demonstrates computational efficiency through its use of inverse STFT for audio reconstruction while achieving competitive performance metrics on speech and music synthesis tasks.

CodeLlama 34B is a large language model developed by Meta that builds upon Llama 2's architecture and is optimized for code generation, understanding, and programming tasks. The model supports multiple programming languages including Python, C++, Java, and JavaScript, with an extended context window of up to 100,000 tokens for handling large codebases. Available in three variants (Base, Python-specialized, and Instruct), it achieved 53.7% accuracy on HumanEval and 56.2% on MBPP benchmarks, demonstrating capabilities in code completion, debugging, and natural language explanations.

Llama 2 7B is a transformer-based language model developed by Meta with 7 billion parameters, trained on 2 trillion tokens with a 4,096-token context length. The model supports text generation in English and 27 other languages, with chat-optimized variants fine-tuned using supervised learning and reinforcement learning from human feedback for dialogue applications.

Llama 2 70B is a 70-billion parameter transformer-based language model developed by Meta, featuring Grouped-Query Attention and a 4096-token context window. Trained on 2 trillion tokens with a September 2022 cutoff, it demonstrates strong performance across language benchmarks including 68.9 on MMLU and 37.5 pass@1 on code generation tasks, while offering both pretrained and chat-optimized variants under Meta's commercial license.

Bark is a transformer-based text-to-audio model that generates multilingual speech, music, and sound effects by converting text directly to audio tokens using EnCodec quantization. The model supports over 13 languages with 100+ speaker presets and can produce nonverbal sounds like laughter through special tokens, operating via a three-stage pipeline from semantic to fine audio tokens.

LLaMA 13B is a transformer-based language model developed by Meta as part of the LLaMA model family, featuring 13 billion parameters and trained on 1.4 trillion tokens from publicly available datasets. The model incorporates architectural optimizations including RMSNorm, SwiGLU activation functions, and rotary positional embeddings, achieving competitive performance with larger models while maintaining efficiency. Released under a noncommercial research license, it demonstrates capabilities across language understanding, reasoning, and code generation benchmarks.

LLaMA 65B is a 65.2 billion parameter transformer-based language model developed by Meta and released in February 2023. The model utilizes architectural optimizations including RMSNorm pre-normalization, SwiGLU activation functions, and rotary positional embeddings. Trained exclusively on 1.4 trillion tokens from publicly available datasets including CommonCrawl, Wikipedia, GitHub, and arXiv, it demonstrates competitive performance across natural language understanding benchmarks while being distributed under a non-commercial research license.

Stable Diffusion 2 is an open-source text-to-image diffusion model developed by Stability AI that generates images at resolutions up to 768×768 pixels using latent diffusion techniques. The model employs an OpenCLIP-ViT/H text encoder and was trained on filtered subsets of the LAION-5B dataset. It includes specialized variants for inpainting, depth-conditioned generation, and 4x upscaling, offering improved capabilities over earlier versions while maintaining open accessibility for research applications.

Stable Diffusion 1.5 is a latent text-to-image diffusion model that generates 512x512 images from text prompts using a U-Net architecture conditioned on CLIP text embeddings within a compressed latent space. Trained on LAION dataset subsets, the model supports text-to-image generation, image-to-image translation, and inpainting tasks, released under the CreativeML OpenRAIL-M license for research and commercial applications.

Stable Diffusion 1.1 is a latent text-to-image diffusion model developed by CompVis, Stability AI, and Runway that generates images from natural language prompts. The model uses a VAE to compress images into latent space, a U-Net for denoising, and a CLIP text encoder for conditioning. Trained on LAION dataset subsets at 512×512 resolution, it supports text-to-image generation, image-to-image translation, and inpainting applications while operating efficiently in compressed latent space.

Devstral is a specialized 23.6 billion parameter language model developed by Mistral AI and All Hands AI, finetuned from Mistral-Small-3.1 for software engineering tasks. The text-only model features a 128,000-token context window and achieves 46.8% on SWE-Bench Verified benchmarks. Released under Apache 2.0 License, it functions as an agentic coding assistant for codebase exploration, multi-file editing, and automated software engineering workflows.

Wan 2.1 T2V 1.3B is an open-source text-to-video generation model developed by Wan-AI, featuring 1.3 billion parameters and utilizing a Flow Matching framework with diffusion transformers. The model supports multilingual text-to-video synthesis in English and Chinese, operates efficiently on consumer GPUs requiring 8.19 GB VRAM, and generates 480P videos with capabilities for image-to-video conversion and text rendering within videos.

Wan 2.1 I2V 14B 720P is a 14-billion parameter image-to-video generation model developed by Wan-AI that converts single images into 720P videos. Built on a unified transformer-based diffusion architecture with a novel 3D causal VAE (Wan-VAE) for spatiotemporal compression, the model supports multilingual text prompts and demonstrates competitive performance in video generation benchmarks while maintaining computational efficiency across various GPU configurations.

Qwen2.5-VL-3B-Instruct is a multimodal large language model developed by Alibaba Cloud featuring 3 billion parameters. The model combines a Vision Transformer encoder with a Qwen2.5-series decoder to process images, videos, and text through dynamic resolution handling and temporal processing capabilities. It supports object detection, OCR, document analysis, video understanding, and computer interface automation, trained on approximately 1.4 trillion tokens across multiple modalities and released under Apache-2.0 license.

Qwen2.5-VL 72B is a 72-billion parameter multimodal generative AI model developed by Alibaba Cloud that integrates vision and language understanding. The model features dynamic resolution processing, temporal video alignment, and architectural enhancements over previous Qwen2-VL versions. It performs object detection, document parsing, video comprehension, OCR across multiple languages, and functions as a visual agent for interactive tasks, trained on over 1.4 trillion tokens.

DeepSeek R1 is a large language model developed by DeepSeek AI that employs a Mixture-of-Experts architecture with 671 billion total parameters and 37 billion activated during inference. The model utilizes reinforcement learning and supervised fine-tuning to enhance reasoning capabilities across mathematics, coding, and logic tasks, achieving competitive performance on benchmarks including 90.8 on MMLU and 97.3 on MATH-500.

DeepSeek V3 is a Mixture-of-Experts language model with 671 billion total parameters and 37 billion activated per token, developed by DeepSeek-AI. The model features Multi-head Latent Attention, auxiliary-loss-free load balancing, and FP8 mixed-precision training. Trained on 14.8 trillion tokens with a 128,000-token context window, it demonstrates competitive performance across reasoning, coding, and mathematical benchmarks while supporting multilingual capabilities and long-context processing.

DeepSeek VL2 Small is a 2.8 billion parameter multimodal vision-language model that uses a Mixture-of-Experts architecture with dynamic tiling for processing high-resolution images. Built on the DeepSeekMoE-16B framework with SigLIP vision encoding, it handles tasks including visual question answering, OCR, document analysis, and visual grounding across multiple languages, achieving competitive performance on benchmarks like DocVQA while maintaining computational efficiency through sparse expert routing.

Phi-4 is a 14-billion parameter decoder-only Transformer language model developed by Microsoft Research that focuses on mathematical reasoning and code generation through curated synthetic data training. The model supports a 16,000-token context window and achieves competitive performance on benchmarks like MMLU (84.8) and HumanEval (82.6) despite its relatively compact size, utilizing supervised fine-tuning and direct preference optimization for alignment.

HunyuanVideo is an open-source video generation model developed by Tencent that supports text-to-video, image-to-video, and controllable video synthesis. The model employs a Transformer-based architecture with a 3D Variational Autoencoder and utilizes flow matching for generating videos at variable resolutions and durations. It features 13 billion parameters and includes capabilities for avatar animation, audio synchronization, and multi-aspect ratio output generation.

CogVideoX 1.5 5B I2V is an image-to-video generation model developed by THUDM using a diffusion transformer architecture with 3D causal variational autoencoder. The model generates temporally coherent videos from input images and text prompts, supporting resolutions up to 1360 pixels and video lengths of 5-10 seconds at 16 fps, trained on 35 million curated video clips.

Qwen2.5-Coder-32B is a 32.5-billion parameter transformer-based language model developed by Alibaba Cloud, specifically designed for programming and code intelligence tasks. The model supports over 92 programming languages and features capabilities in code generation, completion, repair, and reasoning with a 128,000-token context window. Trained on approximately 5.5 trillion tokens of code and instructional data, it demonstrates performance across various coding benchmarks including HumanEval, MBPP, and multilingual programming evaluations.

Llama 3.2 3B is a multilingual instruction-tuned language model developed by Meta with 3 billion parameters and a 128,000-token context window. The model utilizes knowledge distillation from larger Llama variants, Grouped-Query Attention for efficient inference, and advanced quantization techniques optimized for PyTorch's ExecuTorch framework. Supporting eight languages, it targets assistant and agentic applications while enabling deployment in resource-constrained environments.

Qwen 2.5 Math 7B is a 7.62-billion parameter language model developed by Alibaba Cloud that specializes in mathematical reasoning tasks in English and Chinese. The model employs chain-of-thought reasoning and tool-integrated approaches using Python interpreters for computational tasks. It demonstrates improved performance over its predecessor on mathematical benchmarks including MATH, GSM8K, and Chinese mathematics evaluations, achieving 83.6 on MATH using chain-of-thought methods.

Qwen2.5-Coder-7B is a 7.61 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, specialized for code generation and reasoning across 92 programming languages. The model features a 128,000-token context window, supports fill-in-the-middle code completion, and was trained on 5.5 trillion tokens of code and text data, demonstrating competitive performance on coding benchmarks like HumanEval and mathematical reasoning tasks.

Qwen 2.5 14B is a 14.7 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, featuring a 128,000 token context window and support for over 29 languages. The model utilizes advanced architectural components including Grouped Query Attention, RoPE embeddings, and SwiGLU activation, and was pretrained on up to 18 trillion tokens of diverse multilingual data for applications in reasoning, coding, and mathematical tasks.

Qwen 2.5 72B is a 72.7 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, released in September 2024. The model features a 128,000-token context window, supports over 29 languages, and demonstrates strong performance on coding, mathematical reasoning, and knowledge benchmarks. Built with architectural improvements including RoPE and SwiGLU activation functions, it excels at structured data handling and serves as a foundation model for fine-tuning applications.

Command R (08-2024) is a 32-billion parameter generative language model developed by Cohere, featuring a 128,000-token context window and support for 23 languages. The model incorporates Grouped Query Attention for enhanced inference efficiency and specializes in retrieval-augmented generation with citation capabilities, tool use, and multilingual comprehension. It demonstrates improved throughput and reduced latency compared to previous versions while offering configurable safety modes for enterprise applications.

Phi-3.5 Mini Instruct is a 3.8 billion parameter decoder-only Transformer model developed by Microsoft that supports multilingual text generation with a 128,000-token context window. The model demonstrates competitive performance across 22 languages and excels in reasoning, code generation, and long-context tasks, achieving an average benchmark score of 61.4 while maintaining efficient resource utilization.

AuraFlow v0.3 is a 6.8 billion parameter, flow-based text-to-image generative model developed by fal.ai. Built on an optimized DiT architecture with Maximal Update Parametrization, it features enhanced prompt following capabilities through comprehensive recaptioning and prompt enhancement pipelines. The model supports multiple aspect ratios and achieved a GenEval score of 0.703, demonstrating effective text-to-image synthesis across diverse artistic styles and photorealistic outputs.

Stable Fast 3D is a transformer-based generative AI model developed by Stability AI that reconstructs textured 3D mesh assets from single input images in approximately 0.5 seconds. The model predicts comprehensive material properties including albedo, roughness, and metallicity, producing UV-unwrapped meshes suitable for integration into rendering pipelines and interactive applications across gaming, virtual reality, and design workflows.

FLUX.1 [schnell] is a 12-billion parameter text-to-image generation model developed by Black Forest Labs using hybrid diffusion transformer architecture with rectified flow and latent adversarial diffusion distillation. The model generates images from text descriptions in 1-4 diffusion steps, supporting variable resolutions and aspect ratios. Released under Apache 2.0 license, it employs flow matching techniques and parallel attention layers for efficient synthesis.

Mistral Large 2 is a dense transformer-based language model developed by Mistral AI with 123 billion parameters and a 128,000-token context window. The model demonstrates strong performance across multilingual tasks, code generation in 80+ programming languages, mathematical reasoning, and function calling capabilities. It achieves 84% on MMLU, 92% on HumanEval, and 93% on GSM8K benchmarks while maintaining concise output generation.

Mistral NeMo 12B is a transformer-based language model developed collaboratively by Mistral AI and NVIDIA, featuring 12 billion parameters and a 128,000-token context window. The model incorporates grouped query attention, quantization-aware training for FP8 inference, and utilizes the custom Tekken tokenizer for improved multilingual and code compression efficiency. Available in both base and instruction-tuned variants, it demonstrates competitive performance on standard benchmarks while supporting function calling and multilingual capabilities across numerous languages including English, Chinese, Arabic, and various European languages.

Llama 3.1 70B is a transformer-based decoder language model developed by Meta with 70 billion parameters, trained on approximately 15 trillion tokens with a 128K context window. The model supports eight languages and demonstrates competitive performance across benchmarks for reasoning, coding, mathematics, and multilingual tasks. It is available under the Llama 3.1 Community License Agreement for research and commercial applications.

Gemma 2 9B is a decoder-only transformer language model developed by Google with 9.24 billion parameters, trained on 8 trillion tokens of diverse text data including web documents, mathematical content, and code. The model utilizes bfloat16 precision and supports efficient inference through quantization and acceleration techniques, achieving 71.3% on MMLU and 40.2% on HumanEval benchmarks while being suitable for resource-constrained deployment environments.

DeepSeek Coder V2 Lite is an open-source Mixture-of-Experts code language model featuring 16 billion total parameters with 2.4 billion active parameters during inference. The model supports 338 programming languages, processes up to 128,000 tokens of context, and demonstrates competitive performance on code generation benchmarks including 81.1% accuracy on Python HumanEval tasks.

Qwen2-72B is a 72.71 billion parameter Transformer-based language model developed by Alibaba Cloud, featuring Group Query Attention and SwiGLU activation functions. The model demonstrates strong performance across diverse benchmarks including MMLU (84.2), HumanEval (64.6), and GSM8K (89.5), with multilingual capabilities spanning 27 languages and extended context handling up to 128,000 tokens for specialized applications.

Yi 1.5 34B is a 34.4 billion parameter decoder-only Transformer language model developed by 01.AI, featuring Grouped-Query Attention and SwiGLU activations. Trained on 3.1 trillion bilingual tokens, it demonstrates capabilities in reasoning, mathematics, and code generation, with variants supporting up to 200,000 token contexts and multimodal understanding through vision-language extensions.

DeepSeek V2 is a large-scale Mixture-of-Experts language model with 236 billion total parameters, activating only 21 billion per token. It features Multi-head Latent Attention for reduced memory usage and supports context lengths up to 128,000 tokens. Trained on 8.1 trillion tokens with emphasis on English and Chinese data, it demonstrates competitive performance across language understanding, code generation, and mathematical reasoning tasks while achieving significant efficiency improvements over dense models.

Phi-3 Mini Instruct is a 3.8 billion parameter instruction-tuned language model developed by Microsoft using a dense decoder-only Transformer architecture. The model supports a 128,000 token context window and was trained on 4.9 trillion tokens of high-quality data, followed by supervised fine-tuning and direct preference optimization. It demonstrates competitive performance in reasoning, mathematics, and code generation tasks among models under 13 billion parameters, with particular strengths in long-context understanding and structured output generation.

Llama 3 8B is an open-weights transformer-based language model developed by Meta, featuring 8 billion parameters and trained on over 15 trillion tokens. The model utilizes grouped-query attention and a 128,000-token vocabulary, supporting 8,192-token context lengths. Available in both pretrained and instruction-tuned variants, it demonstrates capabilities in text generation, code completion, and conversational tasks across multiple languages.

Command R+ v01 is a 104-billion parameter open-weights language model developed by Cohere, optimized for retrieval-augmented generation, tool use, and multilingual tasks. The model features a 128,000-token context window and specializes in generating outputs with inline citations from retrieved documents. It supports automated tool calling, demonstrates competitive performance across standard benchmarks, and includes efficient tokenization for non-English languages, making it suitable for enterprise applications requiring factual accuracy and transparency.

Command R v01 is a 35-billion-parameter transformer-based language model developed by Cohere, featuring retrieval-augmented generation with explicit citations, tool use capabilities, and multilingual support across ten languages. The model supports a 128,000-token context window and demonstrates performance in enterprise applications, multi-step reasoning tasks, and long-context evaluations, though it requires commercial licensing for enterprise use.

Playground v2.5 Aesthetic is a diffusion-based text-to-image model that generates images at 1024x1024 resolution across multiple aspect ratios. Developed by Playground and released in February 2024, it employs the EDM training framework and human preference alignment techniques to improve color vibrancy, contrast, and human feature rendering compared to its predecessor and other open-source models like Stable Diffusion XL.

Stable Cascade Stage B is an intermediate latent super-resolution component within Stability AI's three-stage text-to-image generation system built on the Würstchen architecture. It operates as a diffusion model that upscales compressed 16×24×24 latents from Stage C to 4×256×256 representations, preserving semantic content while restoring fine details. Available in 700M and 1.5B parameter versions, Stage B enables the system's efficient 42:1 compression ratio and supports extensions like ControlNet and LoRA for enhanced creative workflows.

Stable Video Diffusion XT 1.1 is a latent diffusion model developed by Stability AI that generates 25-frame video sequences at 1024x576 resolution from single input images. The model employs a three-stage training process including image pretraining, video training on curated datasets, and high-resolution finetuning, enabling motion synthesis with configurable camera controls and temporal consistency for image-to-video transformation applications.

Qwen 1.5 72B is a 72-billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team and released in February 2024. The model supports a 32,768-token context window and demonstrates strong multilingual capabilities across 12 languages, achieving competitive performance on benchmarks including MMLU (77.5), C-Eval (84.1), and GSM8K (79.5). It features alignment optimization through Direct Policy Optimization and Proximal Policy Optimization techniques, enabling effective instruction-following and integration with external systems for applications including retrieval-augmented generation and code interpretation.

The SDXL Motion Model is an AnimateDiff-based video generation framework that adds temporal animation capabilities to existing text-to-image diffusion models. Built for compatibility with SDXL at 1024×1024 resolution, it employs a plug-and-play motion module trained on video datasets to generate coherent animated sequences while preserving the visual style of the underlying image model.

Phi-2 is a 2.7 billion parameter Transformer-based language model developed by Microsoft Research and released in December 2023. The model was trained on approximately 1.4 trillion tokens using a "textbook-quality" data approach, incorporating synthetic data from GPT-3.5 and filtered web sources. Phi-2 demonstrates competitive performance in reasoning, language understanding, and code generation tasks compared to larger models in its parameter class.

Mixtral 8x7B is a sparse Mixture of Experts language model developed by Mistral AI and released under the Apache 2.0 license in December 2023. The model uses a decoder-only transformer architecture with eight expert networks per layer, activating only two experts per token, resulting in 12.9 billion active parameters from a total 46.7 billion. It demonstrates competitive performance on benchmarks including MMLU, achieving multilingual capabilities across English, French, German, Spanish, and Italian while maintaining efficient inference speeds.

Playground v2 Aesthetic is a latent diffusion text-to-image model developed by playgroundai that generates 1024x1024 pixel images using dual pre-trained text encoders (OpenCLIP-ViT/G and CLIP-ViT/L). The model achieved a 7.07 FID score on the MJHQ-30K benchmark and demonstrated a 2.5x preference rate over Stable Diffusion XL in user studies, focusing on high-aesthetic image synthesis with strong prompt alignment.

Stable Video Diffusion XT is a generative AI model developed by Stability AI that extends the Stable Diffusion architecture for video synthesis. The model supports image-to-video and text-to-video generation, producing up to 25 frames at resolutions supporting 3-30 fps. Built on a latent video diffusion architecture with over 1.5 billion parameters, SVD-XT incorporates temporal modeling layers and was trained using a three-stage methodology on curated video datasets.

Yi 1 34B is a bilingual transformer-based language model developed by 01.AI, trained on 3 trillion tokens with support for both English and Chinese. The model features a 4,096-token context window and demonstrates competitive performance on multilingual benchmarks including MMLU, CMMLU, and C-Eval, with variants available including extended 200K context and chat-optimized versions released under Apache 2.0 license.

MusicGen is a text-to-music generation model developed by Meta's FAIR team as part of the AudioCraft library. The model uses a two-stage architecture combining EnCodec neural audio compression with a transformer-based autoregressive language model to generate musical audio from textual descriptions or melody inputs. Trained on approximately 20,000 hours of licensed music, MusicGen supports both monophonic and stereophonic outputs and demonstrates competitive performance in objective and subjective evaluations against contemporary music generation models.

Vocos is a neural vocoder developed by GemeloAI that employs a Fourier-based architecture to generate Short-Time Fourier Transform spectral coefficients rather than directly modeling time-domain waveforms. The model supports both mel-spectrogram and neural audio codec token inputs, operates under the MIT license, and demonstrates computational efficiency through its use of inverse STFT for audio reconstruction while achieving competitive performance metrics on speech and music synthesis tasks.

CodeLlama 34B is a large language model developed by Meta that builds upon Llama 2's architecture and is optimized for code generation, understanding, and programming tasks. The model supports multiple programming languages including Python, C++, Java, and JavaScript, with an extended context window of up to 100,000 tokens for handling large codebases. Available in three variants (Base, Python-specialized, and Instruct), it achieved 53.7% accuracy on HumanEval and 56.2% on MBPP benchmarks, demonstrating capabilities in code completion, debugging, and natural language explanations.

Llama 2 7B is a transformer-based language model developed by Meta with 7 billion parameters, trained on 2 trillion tokens with a 4,096-token context length. The model supports text generation in English and 27 other languages, with chat-optimized variants fine-tuned using supervised learning and reinforcement learning from human feedback for dialogue applications.

Llama 2 70B is a 70-billion parameter transformer-based language model developed by Meta, featuring Grouped-Query Attention and a 4096-token context window. Trained on 2 trillion tokens with a September 2022 cutoff, it demonstrates strong performance across language benchmarks including 68.9 on MMLU and 37.5 pass@1 on code generation tasks, while offering both pretrained and chat-optimized variants under Meta's commercial license.

Bark is a transformer-based text-to-audio model that generates multilingual speech, music, and sound effects by converting text directly to audio tokens using EnCodec quantization. The model supports over 13 languages with 100+ speaker presets and can produce nonverbal sounds like laughter through special tokens, operating via a three-stage pipeline from semantic to fine audio tokens.

LLaMA 13B is a transformer-based language model developed by Meta as part of the LLaMA model family, featuring 13 billion parameters and trained on 1.4 trillion tokens from publicly available datasets. The model incorporates architectural optimizations including RMSNorm, SwiGLU activation functions, and rotary positional embeddings, achieving competitive performance with larger models while maintaining efficiency. Released under a noncommercial research license, it demonstrates capabilities across language understanding, reasoning, and code generation benchmarks.

LLaMA 65B is a 65.2 billion parameter transformer-based language model developed by Meta and released in February 2023. The model utilizes architectural optimizations including RMSNorm pre-normalization, SwiGLU activation functions, and rotary positional embeddings. Trained exclusively on 1.4 trillion tokens from publicly available datasets including CommonCrawl, Wikipedia, GitHub, and arXiv, it demonstrates competitive performance across natural language understanding benchmarks while being distributed under a non-commercial research license.

Stable Diffusion 2 is an open-source text-to-image diffusion model developed by Stability AI that generates images at resolutions up to 768×768 pixels using latent diffusion techniques. The model employs an OpenCLIP-ViT/H text encoder and was trained on filtered subsets of the LAION-5B dataset. It includes specialized variants for inpainting, depth-conditioned generation, and 4x upscaling, offering improved capabilities over earlier versions while maintaining open accessibility for research applications.

Stable Diffusion 1.5 is a latent text-to-image diffusion model that generates 512x512 images from text prompts using a U-Net architecture conditioned on CLIP text embeddings within a compressed latent space. Trained on LAION dataset subsets, the model supports text-to-image generation, image-to-image translation, and inpainting tasks, released under the CreativeML OpenRAIL-M license for research and commercial applications.

Stable Diffusion 1.1 is a latent text-to-image diffusion model developed by CompVis, Stability AI, and Runway that generates images from natural language prompts. The model uses a VAE to compress images into latent space, a U-Net for denoising, and a CLIP text encoder for conditioning. Trained on LAION dataset subsets at 512×512 resolution, it supports text-to-image generation, image-to-image translation, and inpainting applications while operating efficiently in compressed latent space.

Devstral is a specialized 23.6 billion parameter language model developed by Mistral AI and All Hands AI, finetuned from Mistral-Small-3.1 for software engineering tasks. The text-only model features a 128,000-token context window and achieves 46.8% on SWE-Bench Verified benchmarks. Released under Apache 2.0 License, it functions as an agentic coding assistant for codebase exploration, multi-file editing, and automated software engineering workflows.

Wan 2.1 T2V 1.3B is an open-source text-to-video generation model developed by Wan-AI, featuring 1.3 billion parameters and utilizing a Flow Matching framework with diffusion transformers. The model supports multilingual text-to-video synthesis in English and Chinese, operates efficiently on consumer GPUs requiring 8.19 GB VRAM, and generates 480P videos with capabilities for image-to-video conversion and text rendering within videos.

Wan 2.1 I2V 14B 720P is a 14-billion parameter image-to-video generation model developed by Wan-AI that converts single images into 720P videos. Built on a unified transformer-based diffusion architecture with a novel 3D causal VAE (Wan-VAE) for spatiotemporal compression, the model supports multilingual text prompts and demonstrates competitive performance in video generation benchmarks while maintaining computational efficiency across various GPU configurations.

Qwen2.5-VL-3B-Instruct is a multimodal large language model developed by Alibaba Cloud featuring 3 billion parameters. The model combines a Vision Transformer encoder with a Qwen2.5-series decoder to process images, videos, and text through dynamic resolution handling and temporal processing capabilities. It supports object detection, OCR, document analysis, video understanding, and computer interface automation, trained on approximately 1.4 trillion tokens across multiple modalities and released under Apache-2.0 license.

Qwen2.5-VL 72B is a 72-billion parameter multimodal generative AI model developed by Alibaba Cloud that integrates vision and language understanding. The model features dynamic resolution processing, temporal video alignment, and architectural enhancements over previous Qwen2-VL versions. It performs object detection, document parsing, video comprehension, OCR across multiple languages, and functions as a visual agent for interactive tasks, trained on over 1.4 trillion tokens.

DeepSeek R1 is a large language model developed by DeepSeek AI that employs a Mixture-of-Experts architecture with 671 billion total parameters and 37 billion activated during inference. The model utilizes reinforcement learning and supervised fine-tuning to enhance reasoning capabilities across mathematics, coding, and logic tasks, achieving competitive performance on benchmarks including 90.8 on MMLU and 97.3 on MATH-500.

DeepSeek V3 is a Mixture-of-Experts language model with 671 billion total parameters and 37 billion activated per token, developed by DeepSeek-AI. The model features Multi-head Latent Attention, auxiliary-loss-free load balancing, and FP8 mixed-precision training. Trained on 14.8 trillion tokens with a 128,000-token context window, it demonstrates competitive performance across reasoning, coding, and mathematical benchmarks while supporting multilingual capabilities and long-context processing.

DeepSeek VL2 Small is a 2.8 billion parameter multimodal vision-language model that uses a Mixture-of-Experts architecture with dynamic tiling for processing high-resolution images. Built on the DeepSeekMoE-16B framework with SigLIP vision encoding, it handles tasks including visual question answering, OCR, document analysis, and visual grounding across multiple languages, achieving competitive performance on benchmarks like DocVQA while maintaining computational efficiency through sparse expert routing.

Phi-4 is a 14-billion parameter decoder-only Transformer language model developed by Microsoft Research that focuses on mathematical reasoning and code generation through curated synthetic data training. The model supports a 16,000-token context window and achieves competitive performance on benchmarks like MMLU (84.8) and HumanEval (82.6) despite its relatively compact size, utilizing supervised fine-tuning and direct preference optimization for alignment.

HunyuanVideo is an open-source video generation model developed by Tencent that supports text-to-video, image-to-video, and controllable video synthesis. The model employs a Transformer-based architecture with a 3D Variational Autoencoder and utilizes flow matching for generating videos at variable resolutions and durations. It features 13 billion parameters and includes capabilities for avatar animation, audio synchronization, and multi-aspect ratio output generation.

CogVideoX 1.5 5B I2V is an image-to-video generation model developed by THUDM using a diffusion transformer architecture with 3D causal variational autoencoder. The model generates temporally coherent videos from input images and text prompts, supporting resolutions up to 1360 pixels and video lengths of 5-10 seconds at 16 fps, trained on 35 million curated video clips.

Qwen2.5-Coder-32B is a 32.5-billion parameter transformer-based language model developed by Alibaba Cloud, specifically designed for programming and code intelligence tasks. The model supports over 92 programming languages and features capabilities in code generation, completion, repair, and reasoning with a 128,000-token context window. Trained on approximately 5.5 trillion tokens of code and instructional data, it demonstrates performance across various coding benchmarks including HumanEval, MBPP, and multilingual programming evaluations.

Llama 3.2 3B is a multilingual instruction-tuned language model developed by Meta with 3 billion parameters and a 128,000-token context window. The model utilizes knowledge distillation from larger Llama variants, Grouped-Query Attention for efficient inference, and advanced quantization techniques optimized for PyTorch's ExecuTorch framework. Supporting eight languages, it targets assistant and agentic applications while enabling deployment in resource-constrained environments.

Qwen 2.5 Math 7B is a 7.62-billion parameter language model developed by Alibaba Cloud that specializes in mathematical reasoning tasks in English and Chinese. The model employs chain-of-thought reasoning and tool-integrated approaches using Python interpreters for computational tasks. It demonstrates improved performance over its predecessor on mathematical benchmarks including MATH, GSM8K, and Chinese mathematics evaluations, achieving 83.6 on MATH using chain-of-thought methods.

Qwen2.5-Coder-7B is a 7.61 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, specialized for code generation and reasoning across 92 programming languages. The model features a 128,000-token context window, supports fill-in-the-middle code completion, and was trained on 5.5 trillion tokens of code and text data, demonstrating competitive performance on coding benchmarks like HumanEval and mathematical reasoning tasks.

Qwen 2.5 14B is a 14.7 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, featuring a 128,000 token context window and support for over 29 languages. The model utilizes advanced architectural components including Grouped Query Attention, RoPE embeddings, and SwiGLU activation, and was pretrained on up to 18 trillion tokens of diverse multilingual data for applications in reasoning, coding, and mathematical tasks.

Qwen 2.5 72B is a 72.7 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, released in September 2024. The model features a 128,000-token context window, supports over 29 languages, and demonstrates strong performance on coding, mathematical reasoning, and knowledge benchmarks. Built with architectural improvements including RoPE and SwiGLU activation functions, it excels at structured data handling and serves as a foundation model for fine-tuning applications.

Command R (08-2024) is a 32-billion parameter generative language model developed by Cohere, featuring a 128,000-token context window and support for 23 languages. The model incorporates Grouped Query Attention for enhanced inference efficiency and specializes in retrieval-augmented generation with citation capabilities, tool use, and multilingual comprehension. It demonstrates improved throughput and reduced latency compared to previous versions while offering configurable safety modes for enterprise applications.

Phi-3.5 Mini Instruct is a 3.8 billion parameter decoder-only Transformer model developed by Microsoft that supports multilingual text generation with a 128,000-token context window. The model demonstrates competitive performance across 22 languages and excels in reasoning, code generation, and long-context tasks, achieving an average benchmark score of 61.4 while maintaining efficient resource utilization.

AuraFlow v0.3 is a 6.8 billion parameter, flow-based text-to-image generative model developed by fal.ai. Built on an optimized DiT architecture with Maximal Update Parametrization, it features enhanced prompt following capabilities through comprehensive recaptioning and prompt enhancement pipelines. The model supports multiple aspect ratios and achieved a GenEval score of 0.703, demonstrating effective text-to-image synthesis across diverse artistic styles and photorealistic outputs.

Stable Fast 3D is a transformer-based generative AI model developed by Stability AI that reconstructs textured 3D mesh assets from single input images in approximately 0.5 seconds. The model predicts comprehensive material properties including albedo, roughness, and metallicity, producing UV-unwrapped meshes suitable for integration into rendering pipelines and interactive applications across gaming, virtual reality, and design workflows.

FLUX.1 [schnell] is a 12-billion parameter text-to-image generation model developed by Black Forest Labs using hybrid diffusion transformer architecture with rectified flow and latent adversarial diffusion distillation. The model generates images from text descriptions in 1-4 diffusion steps, supporting variable resolutions and aspect ratios. Released under Apache 2.0 license, it employs flow matching techniques and parallel attention layers for efficient synthesis.

Mistral Large 2 is a dense transformer-based language model developed by Mistral AI with 123 billion parameters and a 128,000-token context window. The model demonstrates strong performance across multilingual tasks, code generation in 80+ programming languages, mathematical reasoning, and function calling capabilities. It achieves 84% on MMLU, 92% on HumanEval, and 93% on GSM8K benchmarks while maintaining concise output generation.

Mistral NeMo 12B is a transformer-based language model developed collaboratively by Mistral AI and NVIDIA, featuring 12 billion parameters and a 128,000-token context window. The model incorporates grouped query attention, quantization-aware training for FP8 inference, and utilizes the custom Tekken tokenizer for improved multilingual and code compression efficiency. Available in both base and instruction-tuned variants, it demonstrates competitive performance on standard benchmarks while supporting function calling and multilingual capabilities across numerous languages including English, Chinese, Arabic, and various European languages.

Llama 3.1 70B is a transformer-based decoder language model developed by Meta with 70 billion parameters, trained on approximately 15 trillion tokens with a 128K context window. The model supports eight languages and demonstrates competitive performance across benchmarks for reasoning, coding, mathematics, and multilingual tasks. It is available under the Llama 3.1 Community License Agreement for research and commercial applications.

Gemma 2 9B is a decoder-only transformer language model developed by Google with 9.24 billion parameters, trained on 8 trillion tokens of diverse text data including web documents, mathematical content, and code. The model utilizes bfloat16 precision and supports efficient inference through quantization and acceleration techniques, achieving 71.3% on MMLU and 40.2% on HumanEval benchmarks while being suitable for resource-constrained deployment environments.

DeepSeek Coder V2 Lite is an open-source Mixture-of-Experts code language model featuring 16 billion total parameters with 2.4 billion active parameters during inference. The model supports 338 programming languages, processes up to 128,000 tokens of context, and demonstrates competitive performance on code generation benchmarks including 81.1% accuracy on Python HumanEval tasks.

Qwen2-72B is a 72.71 billion parameter Transformer-based language model developed by Alibaba Cloud, featuring Group Query Attention and SwiGLU activation functions. The model demonstrates strong performance across diverse benchmarks including MMLU (84.2), HumanEval (64.6), and GSM8K (89.5), with multilingual capabilities spanning 27 languages and extended context handling up to 128,000 tokens for specialized applications.

Yi 1.5 34B is a 34.4 billion parameter decoder-only Transformer language model developed by 01.AI, featuring Grouped-Query Attention and SwiGLU activations. Trained on 3.1 trillion bilingual tokens, it demonstrates capabilities in reasoning, mathematics, and code generation, with variants supporting up to 200,000 token contexts and multimodal understanding through vision-language extensions.

DeepSeek V2 is a large-scale Mixture-of-Experts language model with 236 billion total parameters, activating only 21 billion per token. It features Multi-head Latent Attention for reduced memory usage and supports context lengths up to 128,000 tokens. Trained on 8.1 trillion tokens with emphasis on English and Chinese data, it demonstrates competitive performance across language understanding, code generation, and mathematical reasoning tasks while achieving significant efficiency improvements over dense models.

Phi-3 Mini Instruct is a 3.8 billion parameter instruction-tuned language model developed by Microsoft using a dense decoder-only Transformer architecture. The model supports a 128,000 token context window and was trained on 4.9 trillion tokens of high-quality data, followed by supervised fine-tuning and direct preference optimization. It demonstrates competitive performance in reasoning, mathematics, and code generation tasks among models under 13 billion parameters, with particular strengths in long-context understanding and structured output generation.

Llama 3 8B is an open-weights transformer-based language model developed by Meta, featuring 8 billion parameters and trained on over 15 trillion tokens. The model utilizes grouped-query attention and a 128,000-token vocabulary, supporting 8,192-token context lengths. Available in both pretrained and instruction-tuned variants, it demonstrates capabilities in text generation, code completion, and conversational tasks across multiple languages.

Command R+ v01 is a 104-billion parameter open-weights language model developed by Cohere, optimized for retrieval-augmented generation, tool use, and multilingual tasks. The model features a 128,000-token context window and specializes in generating outputs with inline citations from retrieved documents. It supports automated tool calling, demonstrates competitive performance across standard benchmarks, and includes efficient tokenization for non-English languages, making it suitable for enterprise applications requiring factual accuracy and transparency.

Command R v01 is a 35-billion-parameter transformer-based language model developed by Cohere, featuring retrieval-augmented generation with explicit citations, tool use capabilities, and multilingual support across ten languages. The model supports a 128,000-token context window and demonstrates performance in enterprise applications, multi-step reasoning tasks, and long-context evaluations, though it requires commercial licensing for enterprise use.

Playground v2.5 Aesthetic is a diffusion-based text-to-image model that generates images at 1024x1024 resolution across multiple aspect ratios. Developed by Playground and released in February 2024, it employs the EDM training framework and human preference alignment techniques to improve color vibrancy, contrast, and human feature rendering compared to its predecessor and other open-source models like Stable Diffusion XL.

Stable Cascade Stage B is an intermediate latent super-resolution component within Stability AI's three-stage text-to-image generation system built on the Würstchen architecture. It operates as a diffusion model that upscales compressed 16×24×24 latents from Stage C to 4×256×256 representations, preserving semantic content while restoring fine details. Available in 700M and 1.5B parameter versions, Stage B enables the system's efficient 42:1 compression ratio and supports extensions like ControlNet and LoRA for enhanced creative workflows.

Stable Video Diffusion XT 1.1 is a latent diffusion model developed by Stability AI that generates 25-frame video sequences at 1024x576 resolution from single input images. The model employs a three-stage training process including image pretraining, video training on curated datasets, and high-resolution finetuning, enabling motion synthesis with configurable camera controls and temporal consistency for image-to-video transformation applications.

Qwen 1.5 72B is a 72-billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team and released in February 2024. The model supports a 32,768-token context window and demonstrates strong multilingual capabilities across 12 languages, achieving competitive performance on benchmarks including MMLU (77.5), C-Eval (84.1), and GSM8K (79.5). It features alignment optimization through Direct Policy Optimization and Proximal Policy Optimization techniques, enabling effective instruction-following and integration with external systems for applications including retrieval-augmented generation and code interpretation.

The SDXL Motion Model is an AnimateDiff-based video generation framework that adds temporal animation capabilities to existing text-to-image diffusion models. Built for compatibility with SDXL at 1024×1024 resolution, it employs a plug-and-play motion module trained on video datasets to generate coherent animated sequences while preserving the visual style of the underlying image model.

Phi-2 is a 2.7 billion parameter Transformer-based language model developed by Microsoft Research and released in December 2023. The model was trained on approximately 1.4 trillion tokens using a "textbook-quality" data approach, incorporating synthetic data from GPT-3.5 and filtered web sources. Phi-2 demonstrates competitive performance in reasoning, language understanding, and code generation tasks compared to larger models in its parameter class.

Mixtral 8x7B is a sparse Mixture of Experts language model developed by Mistral AI and released under the Apache 2.0 license in December 2023. The model uses a decoder-only transformer architecture with eight expert networks per layer, activating only two experts per token, resulting in 12.9 billion active parameters from a total 46.7 billion. It demonstrates competitive performance on benchmarks including MMLU, achieving multilingual capabilities across English, French, German, Spanish, and Italian while maintaining efficient inference speeds.

Playground v2 Aesthetic is a latent diffusion text-to-image model developed by playgroundai that generates 1024x1024 pixel images using dual pre-trained text encoders (OpenCLIP-ViT/G and CLIP-ViT/L). The model achieved a 7.07 FID score on the MJHQ-30K benchmark and demonstrated a 2.5x preference rate over Stable Diffusion XL in user studies, focusing on high-aesthetic image synthesis with strong prompt alignment.

Stable Video Diffusion XT is a generative AI model developed by Stability AI that extends the Stable Diffusion architecture for video synthesis. The model supports image-to-video and text-to-video generation, producing up to 25 frames at resolutions supporting 3-30 fps. Built on a latent video diffusion architecture with over 1.5 billion parameters, SVD-XT incorporates temporal modeling layers and was trained using a three-stage methodology on curated video datasets.

Yi 1 34B is a bilingual transformer-based language model developed by 01.AI, trained on 3 trillion tokens with support for both English and Chinese. The model features a 4,096-token context window and demonstrates competitive performance on multilingual benchmarks including MMLU, CMMLU, and C-Eval, with variants available including extended 200K context and chat-optimized versions released under Apache 2.0 license.

MusicGen is a text-to-music generation model developed by Meta's FAIR team as part of the AudioCraft library. The model uses a two-stage architecture combining EnCodec neural audio compression with a transformer-based autoregressive language model to generate musical audio from textual descriptions or melody inputs. Trained on approximately 20,000 hours of licensed music, MusicGen supports both monophonic and stereophonic outputs and demonstrates competitive performance in objective and subjective evaluations against contemporary music generation models.

Vocos is a neural vocoder developed by GemeloAI that employs a Fourier-based architecture to generate Short-Time Fourier Transform spectral coefficients rather than directly modeling time-domain waveforms. The model supports both mel-spectrogram and neural audio codec token inputs, operates under the MIT license, and demonstrates computational efficiency through its use of inverse STFT for audio reconstruction while achieving competitive performance metrics on speech and music synthesis tasks.

CodeLlama 34B is a large language model developed by Meta that builds upon Llama 2's architecture and is optimized for code generation, understanding, and programming tasks. The model supports multiple programming languages including Python, C++, Java, and JavaScript, with an extended context window of up to 100,000 tokens for handling large codebases. Available in three variants (Base, Python-specialized, and Instruct), it achieved 53.7% accuracy on HumanEval and 56.2% on MBPP benchmarks, demonstrating capabilities in code completion, debugging, and natural language explanations.

Llama 2 7B is a transformer-based language model developed by Meta with 7 billion parameters, trained on 2 trillion tokens with a 4,096-token context length. The model supports text generation in English and 27 other languages, with chat-optimized variants fine-tuned using supervised learning and reinforcement learning from human feedback for dialogue applications.

Llama 2 70B is a 70-billion parameter transformer-based language model developed by Meta, featuring Grouped-Query Attention and a 4096-token context window. Trained on 2 trillion tokens with a September 2022 cutoff, it demonstrates strong performance across language benchmarks including 68.9 on MMLU and 37.5 pass@1 on code generation tasks, while offering both pretrained and chat-optimized variants under Meta's commercial license.

Bark is a transformer-based text-to-audio model that generates multilingual speech, music, and sound effects by converting text directly to audio tokens using EnCodec quantization. The model supports over 13 languages with 100+ speaker presets and can produce nonverbal sounds like laughter through special tokens, operating via a three-stage pipeline from semantic to fine audio tokens.

LLaMA 13B is a transformer-based language model developed by Meta as part of the LLaMA model family, featuring 13 billion parameters and trained on 1.4 trillion tokens from publicly available datasets. The model incorporates architectural optimizations including RMSNorm, SwiGLU activation functions, and rotary positional embeddings, achieving competitive performance with larger models while maintaining efficiency. Released under a noncommercial research license, it demonstrates capabilities across language understanding, reasoning, and code generation benchmarks.

LLaMA 65B is a 65.2 billion parameter transformer-based language model developed by Meta and released in February 2023. The model utilizes architectural optimizations including RMSNorm pre-normalization, SwiGLU activation functions, and rotary positional embeddings. Trained exclusively on 1.4 trillion tokens from publicly available datasets including CommonCrawl, Wikipedia, GitHub, and arXiv, it demonstrates competitive performance across natural language understanding benchmarks while being distributed under a non-commercial research license.

Stable Diffusion 2 is an open-source text-to-image diffusion model developed by Stability AI that generates images at resolutions up to 768×768 pixels using latent diffusion techniques. The model employs an OpenCLIP-ViT/H text encoder and was trained on filtered subsets of the LAION-5B dataset. It includes specialized variants for inpainting, depth-conditioned generation, and 4x upscaling, offering improved capabilities over earlier versions while maintaining open accessibility for research applications.

Stable Diffusion 1.5 is a latent text-to-image diffusion model that generates 512x512 images from text prompts using a U-Net architecture conditioned on CLIP text embeddings within a compressed latent space. Trained on LAION dataset subsets, the model supports text-to-image generation, image-to-image translation, and inpainting tasks, released under the CreativeML OpenRAIL-M license for research and commercial applications.

Stable Diffusion 1.1 is a latent text-to-image diffusion model developed by CompVis, Stability AI, and Runway that generates images from natural language prompts. The model uses a VAE to compress images into latent space, a U-Net for denoising, and a CLIP text encoder for conditioning. Trained on LAION dataset subsets at 512×512 resolution, it supports text-to-image generation, image-to-image translation, and inpainting applications while operating efficiently in compressed latent space.

Devstral is a specialized 23.6 billion parameter language model developed by Mistral AI and All Hands AI, finetuned from Mistral-Small-3.1 for software engineering tasks. The text-only model features a 128,000-token context window and achieves 46.8% on SWE-Bench Verified benchmarks. Released under Apache 2.0 License, it functions as an agentic coding assistant for codebase exploration, multi-file editing, and automated software engineering workflows.

Wan 2.1 T2V 1.3B is an open-source text-to-video generation model developed by Wan-AI, featuring 1.3 billion parameters and utilizing a Flow Matching framework with diffusion transformers. The model supports multilingual text-to-video synthesis in English and Chinese, operates efficiently on consumer GPUs requiring 8.19 GB VRAM, and generates 480P videos with capabilities for image-to-video conversion and text rendering within videos.

Wan 2.1 I2V 14B 720P is a 14-billion parameter image-to-video generation model developed by Wan-AI that converts single images into 720P videos. Built on a unified transformer-based diffusion architecture with a novel 3D causal VAE (Wan-VAE) for spatiotemporal compression, the model supports multilingual text prompts and demonstrates competitive performance in video generation benchmarks while maintaining computational efficiency across various GPU configurations.

Qwen2.5-VL-3B-Instruct is a multimodal large language model developed by Alibaba Cloud featuring 3 billion parameters. The model combines a Vision Transformer encoder with a Qwen2.5-series decoder to process images, videos, and text through dynamic resolution handling and temporal processing capabilities. It supports object detection, OCR, document analysis, video understanding, and computer interface automation, trained on approximately 1.4 trillion tokens across multiple modalities and released under Apache-2.0 license.

Qwen2.5-VL 72B is a 72-billion parameter multimodal generative AI model developed by Alibaba Cloud that integrates vision and language understanding. The model features dynamic resolution processing, temporal video alignment, and architectural enhancements over previous Qwen2-VL versions. It performs object detection, document parsing, video comprehension, OCR across multiple languages, and functions as a visual agent for interactive tasks, trained on over 1.4 trillion tokens.

DeepSeek R1 is a large language model developed by DeepSeek AI that employs a Mixture-of-Experts architecture with 671 billion total parameters and 37 billion activated during inference. The model utilizes reinforcement learning and supervised fine-tuning to enhance reasoning capabilities across mathematics, coding, and logic tasks, achieving competitive performance on benchmarks including 90.8 on MMLU and 97.3 on MATH-500.

DeepSeek V3 is a Mixture-of-Experts language model with 671 billion total parameters and 37 billion activated per token, developed by DeepSeek-AI. The model features Multi-head Latent Attention, auxiliary-loss-free load balancing, and FP8 mixed-precision training. Trained on 14.8 trillion tokens with a 128,000-token context window, it demonstrates competitive performance across reasoning, coding, and mathematical benchmarks while supporting multilingual capabilities and long-context processing.

DeepSeek VL2 Small is a 2.8 billion parameter multimodal vision-language model that uses a Mixture-of-Experts architecture with dynamic tiling for processing high-resolution images. Built on the DeepSeekMoE-16B framework with SigLIP vision encoding, it handles tasks including visual question answering, OCR, document analysis, and visual grounding across multiple languages, achieving competitive performance on benchmarks like DocVQA while maintaining computational efficiency through sparse expert routing.

Phi-4 is a 14-billion parameter decoder-only Transformer language model developed by Microsoft Research that focuses on mathematical reasoning and code generation through curated synthetic data training. The model supports a 16,000-token context window and achieves competitive performance on benchmarks like MMLU (84.8) and HumanEval (82.6) despite its relatively compact size, utilizing supervised fine-tuning and direct preference optimization for alignment.

HunyuanVideo is an open-source video generation model developed by Tencent that supports text-to-video, image-to-video, and controllable video synthesis. The model employs a Transformer-based architecture with a 3D Variational Autoencoder and utilizes flow matching for generating videos at variable resolutions and durations. It features 13 billion parameters and includes capabilities for avatar animation, audio synchronization, and multi-aspect ratio output generation.

CogVideoX 1.5 5B I2V is an image-to-video generation model developed by THUDM using a diffusion transformer architecture with 3D causal variational autoencoder. The model generates temporally coherent videos from input images and text prompts, supporting resolutions up to 1360 pixels and video lengths of 5-10 seconds at 16 fps, trained on 35 million curated video clips.

Qwen2.5-Coder-32B is a 32.5-billion parameter transformer-based language model developed by Alibaba Cloud, specifically designed for programming and code intelligence tasks. The model supports over 92 programming languages and features capabilities in code generation, completion, repair, and reasoning with a 128,000-token context window. Trained on approximately 5.5 trillion tokens of code and instructional data, it demonstrates performance across various coding benchmarks including HumanEval, MBPP, and multilingual programming evaluations.

Llama 3.2 3B is a multilingual instruction-tuned language model developed by Meta with 3 billion parameters and a 128,000-token context window. The model utilizes knowledge distillation from larger Llama variants, Grouped-Query Attention for efficient inference, and advanced quantization techniques optimized for PyTorch's ExecuTorch framework. Supporting eight languages, it targets assistant and agentic applications while enabling deployment in resource-constrained environments.

Qwen 2.5 Math 7B is a 7.62-billion parameter language model developed by Alibaba Cloud that specializes in mathematical reasoning tasks in English and Chinese. The model employs chain-of-thought reasoning and tool-integrated approaches using Python interpreters for computational tasks. It demonstrates improved performance over its predecessor on mathematical benchmarks including MATH, GSM8K, and Chinese mathematics evaluations, achieving 83.6 on MATH using chain-of-thought methods.

Qwen2.5-Coder-7B is a 7.61 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, specialized for code generation and reasoning across 92 programming languages. The model features a 128,000-token context window, supports fill-in-the-middle code completion, and was trained on 5.5 trillion tokens of code and text data, demonstrating competitive performance on coding benchmarks like HumanEval and mathematical reasoning tasks.

Qwen 2.5 14B is a 14.7 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, featuring a 128,000 token context window and support for over 29 languages. The model utilizes advanced architectural components including Grouped Query Attention, RoPE embeddings, and SwiGLU activation, and was pretrained on up to 18 trillion tokens of diverse multilingual data for applications in reasoning, coding, and mathematical tasks.

Qwen 2.5 72B is a 72.7 billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team, released in September 2024. The model features a 128,000-token context window, supports over 29 languages, and demonstrates strong performance on coding, mathematical reasoning, and knowledge benchmarks. Built with architectural improvements including RoPE and SwiGLU activation functions, it excels at structured data handling and serves as a foundation model for fine-tuning applications.

Command R (08-2024) is a 32-billion parameter generative language model developed by Cohere, featuring a 128,000-token context window and support for 23 languages. The model incorporates Grouped Query Attention for enhanced inference efficiency and specializes in retrieval-augmented generation with citation capabilities, tool use, and multilingual comprehension. It demonstrates improved throughput and reduced latency compared to previous versions while offering configurable safety modes for enterprise applications.

Phi-3.5 Mini Instruct is a 3.8 billion parameter decoder-only Transformer model developed by Microsoft that supports multilingual text generation with a 128,000-token context window. The model demonstrates competitive performance across 22 languages and excels in reasoning, code generation, and long-context tasks, achieving an average benchmark score of 61.4 while maintaining efficient resource utilization.

AuraFlow v0.3 is a 6.8 billion parameter, flow-based text-to-image generative model developed by fal.ai. Built on an optimized DiT architecture with Maximal Update Parametrization, it features enhanced prompt following capabilities through comprehensive recaptioning and prompt enhancement pipelines. The model supports multiple aspect ratios and achieved a GenEval score of 0.703, demonstrating effective text-to-image synthesis across diverse artistic styles and photorealistic outputs.

Stable Fast 3D is a transformer-based generative AI model developed by Stability AI that reconstructs textured 3D mesh assets from single input images in approximately 0.5 seconds. The model predicts comprehensive material properties including albedo, roughness, and metallicity, producing UV-unwrapped meshes suitable for integration into rendering pipelines and interactive applications across gaming, virtual reality, and design workflows.

FLUX.1 [schnell] is a 12-billion parameter text-to-image generation model developed by Black Forest Labs using hybrid diffusion transformer architecture with rectified flow and latent adversarial diffusion distillation. The model generates images from text descriptions in 1-4 diffusion steps, supporting variable resolutions and aspect ratios. Released under Apache 2.0 license, it employs flow matching techniques and parallel attention layers for efficient synthesis.

Mistral Large 2 is a dense transformer-based language model developed by Mistral AI with 123 billion parameters and a 128,000-token context window. The model demonstrates strong performance across multilingual tasks, code generation in 80+ programming languages, mathematical reasoning, and function calling capabilities. It achieves 84% on MMLU, 92% on HumanEval, and 93% on GSM8K benchmarks while maintaining concise output generation.

Mistral NeMo 12B is a transformer-based language model developed collaboratively by Mistral AI and NVIDIA, featuring 12 billion parameters and a 128,000-token context window. The model incorporates grouped query attention, quantization-aware training for FP8 inference, and utilizes the custom Tekken tokenizer for improved multilingual and code compression efficiency. Available in both base and instruction-tuned variants, it demonstrates competitive performance on standard benchmarks while supporting function calling and multilingual capabilities across numerous languages including English, Chinese, Arabic, and various European languages.

Llama 3.1 70B is a transformer-based decoder language model developed by Meta with 70 billion parameters, trained on approximately 15 trillion tokens with a 128K context window. The model supports eight languages and demonstrates competitive performance across benchmarks for reasoning, coding, mathematics, and multilingual tasks. It is available under the Llama 3.1 Community License Agreement for research and commercial applications.

Gemma 2 9B is a decoder-only transformer language model developed by Google with 9.24 billion parameters, trained on 8 trillion tokens of diverse text data including web documents, mathematical content, and code. The model utilizes bfloat16 precision and supports efficient inference through quantization and acceleration techniques, achieving 71.3% on MMLU and 40.2% on HumanEval benchmarks while being suitable for resource-constrained deployment environments.

DeepSeek Coder V2 Lite is an open-source Mixture-of-Experts code language model featuring 16 billion total parameters with 2.4 billion active parameters during inference. The model supports 338 programming languages, processes up to 128,000 tokens of context, and demonstrates competitive performance on code generation benchmarks including 81.1% accuracy on Python HumanEval tasks.

Qwen2-72B is a 72.71 billion parameter Transformer-based language model developed by Alibaba Cloud, featuring Group Query Attention and SwiGLU activation functions. The model demonstrates strong performance across diverse benchmarks including MMLU (84.2), HumanEval (64.6), and GSM8K (89.5), with multilingual capabilities spanning 27 languages and extended context handling up to 128,000 tokens for specialized applications.

Yi 1.5 34B is a 34.4 billion parameter decoder-only Transformer language model developed by 01.AI, featuring Grouped-Query Attention and SwiGLU activations. Trained on 3.1 trillion bilingual tokens, it demonstrates capabilities in reasoning, mathematics, and code generation, with variants supporting up to 200,000 token contexts and multimodal understanding through vision-language extensions.

DeepSeek V2 is a large-scale Mixture-of-Experts language model with 236 billion total parameters, activating only 21 billion per token. It features Multi-head Latent Attention for reduced memory usage and supports context lengths up to 128,000 tokens. Trained on 8.1 trillion tokens with emphasis on English and Chinese data, it demonstrates competitive performance across language understanding, code generation, and mathematical reasoning tasks while achieving significant efficiency improvements over dense models.

Phi-3 Mini Instruct is a 3.8 billion parameter instruction-tuned language model developed by Microsoft using a dense decoder-only Transformer architecture. The model supports a 128,000 token context window and was trained on 4.9 trillion tokens of high-quality data, followed by supervised fine-tuning and direct preference optimization. It demonstrates competitive performance in reasoning, mathematics, and code generation tasks among models under 13 billion parameters, with particular strengths in long-context understanding and structured output generation.

Llama 3 8B is an open-weights transformer-based language model developed by Meta, featuring 8 billion parameters and trained on over 15 trillion tokens. The model utilizes grouped-query attention and a 128,000-token vocabulary, supporting 8,192-token context lengths. Available in both pretrained and instruction-tuned variants, it demonstrates capabilities in text generation, code completion, and conversational tasks across multiple languages.

Command R+ v01 is a 104-billion parameter open-weights language model developed by Cohere, optimized for retrieval-augmented generation, tool use, and multilingual tasks. The model features a 128,000-token context window and specializes in generating outputs with inline citations from retrieved documents. It supports automated tool calling, demonstrates competitive performance across standard benchmarks, and includes efficient tokenization for non-English languages, making it suitable for enterprise applications requiring factual accuracy and transparency.

Command R v01 is a 35-billion-parameter transformer-based language model developed by Cohere, featuring retrieval-augmented generation with explicit citations, tool use capabilities, and multilingual support across ten languages. The model supports a 128,000-token context window and demonstrates performance in enterprise applications, multi-step reasoning tasks, and long-context evaluations, though it requires commercial licensing for enterprise use.

Playground v2.5 Aesthetic is a diffusion-based text-to-image model that generates images at 1024x1024 resolution across multiple aspect ratios. Developed by Playground and released in February 2024, it employs the EDM training framework and human preference alignment techniques to improve color vibrancy, contrast, and human feature rendering compared to its predecessor and other open-source models like Stable Diffusion XL.

Stable Cascade Stage B is an intermediate latent super-resolution component within Stability AI's three-stage text-to-image generation system built on the Würstchen architecture. It operates as a diffusion model that upscales compressed 16×24×24 latents from Stage C to 4×256×256 representations, preserving semantic content while restoring fine details. Available in 700M and 1.5B parameter versions, Stage B enables the system's efficient 42:1 compression ratio and supports extensions like ControlNet and LoRA for enhanced creative workflows.

Stable Video Diffusion XT 1.1 is a latent diffusion model developed by Stability AI that generates 25-frame video sequences at 1024x576 resolution from single input images. The model employs a three-stage training process including image pretraining, video training on curated datasets, and high-resolution finetuning, enabling motion synthesis with configurable camera controls and temporal consistency for image-to-video transformation applications.

Qwen 1.5 72B is a 72-billion parameter transformer-based language model developed by Alibaba Cloud's Qwen Team and released in February 2024. The model supports a 32,768-token context window and demonstrates strong multilingual capabilities across 12 languages, achieving competitive performance on benchmarks including MMLU (77.5), C-Eval (84.1), and GSM8K (79.5). It features alignment optimization through Direct Policy Optimization and Proximal Policy Optimization techniques, enabling effective instruction-following and integration with external systems for applications including retrieval-augmented generation and code interpretation.

The SDXL Motion Model is an AnimateDiff-based video generation framework that adds temporal animation capabilities to existing text-to-image diffusion models. Built for compatibility with SDXL at 1024×1024 resolution, it employs a plug-and-play motion module trained on video datasets to generate coherent animated sequences while preserving the visual style of the underlying image model.

Phi-2 is a 2.7 billion parameter Transformer-based language model developed by Microsoft Research and released in December 2023. The model was trained on approximately 1.4 trillion tokens using a "textbook-quality" data approach, incorporating synthetic data from GPT-3.5 and filtered web sources. Phi-2 demonstrates competitive performance in reasoning, language understanding, and code generation tasks compared to larger models in its parameter class.

Mixtral 8x7B is a sparse Mixture of Experts language model developed by Mistral AI and released under the Apache 2.0 license in December 2023. The model uses a decoder-only transformer architecture with eight expert networks per layer, activating only two experts per token, resulting in 12.9 billion active parameters from a total 46.7 billion. It demonstrates competitive performance on benchmarks including MMLU, achieving multilingual capabilities across English, French, German, Spanish, and Italian while maintaining efficient inference speeds.

Playground v2 Aesthetic is a latent diffusion text-to-image model developed by playgroundai that generates 1024x1024 pixel images using dual pre-trained text encoders (OpenCLIP-ViT/G and CLIP-ViT/L). The model achieved a 7.07 FID score on the MJHQ-30K benchmark and demonstrated a 2.5x preference rate over Stable Diffusion XL in user studies, focusing on high-aesthetic image synthesis with strong prompt alignment.

Stable Video Diffusion XT is a generative AI model developed by Stability AI that extends the Stable Diffusion architecture for video synthesis. The model supports image-to-video and text-to-video generation, producing up to 25 frames at resolutions supporting 3-30 fps. Built on a latent video diffusion architecture with over 1.5 billion parameters, SVD-XT incorporates temporal modeling layers and was trained using a three-stage methodology on curated video datasets.

Yi 1 34B is a bilingual transformer-based language model developed by 01.AI, trained on 3 trillion tokens with support for both English and Chinese. The model features a 4,096-token context window and demonstrates competitive performance on multilingual benchmarks including MMLU, CMMLU, and C-Eval, with variants available including extended 200K context and chat-optimized versions released under Apache 2.0 license.

MusicGen is a text-to-music generation model developed by Meta's FAIR team as part of the AudioCraft library. The model uses a two-stage architecture combining EnCodec neural audio compression with a transformer-based autoregressive language model to generate musical audio from textual descriptions or melody inputs. Trained on approximately 20,000 hours of licensed music, MusicGen supports both monophonic and stereophonic outputs and demonstrates competitive performance in objective and subjective evaluations against contemporary music generation models.

Vocos is a neural vocoder developed by GemeloAI that employs a Fourier-based architecture to generate Short-Time Fourier Transform spectral coefficients rather than directly modeling time-domain waveforms. The model supports both mel-spectrogram and neural audio codec token inputs, operates under the MIT license, and demonstrates computational efficiency through its use of inverse STFT for audio reconstruction while achieving competitive performance metrics on speech and music synthesis tasks.

CodeLlama 34B is a large language model developed by Meta that builds upon Llama 2's architecture and is optimized for code generation, understanding, and programming tasks. The model supports multiple programming languages including Python, C++, Java, and JavaScript, with an extended context window of up to 100,000 tokens for handling large codebases. Available in three variants (Base, Python-specialized, and Instruct), it achieved 53.7% accuracy on HumanEval and 56.2% on MBPP benchmarks, demonstrating capabilities in code completion, debugging, and natural language explanations.

Llama 2 7B is a transformer-based language model developed by Meta with 7 billion parameters, trained on 2 trillion tokens with a 4,096-token context length. The model supports text generation in English and 27 other languages, with chat-optimized variants fine-tuned using supervised learning and reinforcement learning from human feedback for dialogue applications.

Llama 2 70B is a 70-billion parameter transformer-based language model developed by Meta, featuring Grouped-Query Attention and a 4096-token context window. Trained on 2 trillion tokens with a September 2022 cutoff, it demonstrates strong performance across language benchmarks including 68.9 on MMLU and 37.5 pass@1 on code generation tasks, while offering both pretrained and chat-optimized variants under Meta's commercial license.

Bark is a transformer-based text-to-audio model that generates multilingual speech, music, and sound effects by converting text directly to audio tokens using EnCodec quantization. The model supports over 13 languages with 100+ speaker presets and can produce nonverbal sounds like laughter through special tokens, operating via a three-stage pipeline from semantic to fine audio tokens.

LLaMA 13B is a transformer-based language model developed by Meta as part of the LLaMA model family, featuring 13 billion parameters and trained on 1.4 trillion tokens from publicly available datasets. The model incorporates architectural optimizations including RMSNorm, SwiGLU activation functions, and rotary positional embeddings, achieving competitive performance with larger models while maintaining efficiency. Released under a noncommercial research license, it demonstrates capabilities across language understanding, reasoning, and code generation benchmarks.

LLaMA 65B is a 65.2 billion parameter transformer-based language model developed by Meta and released in February 2023. The model utilizes architectural optimizations including RMSNorm pre-normalization, SwiGLU activation functions, and rotary positional embeddings. Trained exclusively on 1.4 trillion tokens from publicly available datasets including CommonCrawl, Wikipedia, GitHub, and arXiv, it demonstrates competitive performance across natural language understanding benchmarks while being distributed under a non-commercial research license.

Stable Diffusion 2 is an open-source text-to-image diffusion model developed by Stability AI that generates images at resolutions up to 768×768 pixels using latent diffusion techniques. The model employs an OpenCLIP-ViT/H text encoder and was trained on filtered subsets of the LAION-5B dataset. It includes specialized variants for inpainting, depth-conditioned generation, and 4x upscaling, offering improved capabilities over earlier versions while maintaining open accessibility for research applications.

Stable Diffusion 1.5 is a latent text-to-image diffusion model that generates 512x512 images from text prompts using a U-Net architecture conditioned on CLIP text embeddings within a compressed latent space. Trained on LAION dataset subsets, the model supports text-to-image generation, image-to-image translation, and inpainting tasks, released under the CreativeML OpenRAIL-M license for research and commercial applications.

Stable Diffusion 1.1 is a latent text-to-image diffusion model developed by CompVis, Stability AI, and Runway that generates images from natural language prompts. The model uses a VAE to compress images into latent space, a U-Net for denoising, and a CLIP text encoder for conditioning. Trained on LAION dataset subsets at 512×512 resolution, it supports text-to-image generation, image-to-image translation, and inpainting applications while operating efficiently in compressed latent space.