Browse Models
Launch a dedicated cloud GPU server running Laboratory OS to download and run Gemma 2 27B using any compatible app or framework.
Must be used with a compatible app, notebook, or codebase. May run slowly, or not work at all, depending on local system resources, particularly GPU(s) and available VRAM.
Open WebUI is an open-source, self-hosted web interface with a polished, ChatGPT-like user experience for interacting with LLMs. Integrates seamlessly with local Ollama installation.
The most full-featured web interface for experimenting with open source Large Language Models. Featuring a wide range of configurable settings, inference engines, and plugins.
Gemma 2 27B is a decoder-only transformer language model developed by Google, trained on 13 trillion tokens using JAX and ML Pathways on TPU hardware. The model achieves 75.2% accuracy on MMLU benchmarks and 51.8% on HumanEval programming tasks. It supports various text generation applications including content creation, dialogue systems, and code assistance, with openly accessible weights distributed under Google's responsible use policies.
Gemma 2 27B is a decoder-only generative language model developed by Google and is part of the Gemma model family. Designed primarily for English-language text-to-text tasks, the model is distributed with openly accessible pre-trained and instruction-tuned weights. The release of Gemma 2 27B is based on the research and technology used to create the Gemini models, integrating developments in large-scale neural network design and model training.
The architecture of Gemma 2 27B employs a transformer-based, decoder-only configuration. The model was trained using Google's Tensor Processing Unit hardware, facilitating large-scale computation. The training stack combines the JAX numerical computing library and ML Pathways, which streamlines development and enables the orchestration of large workloads within a unified programming environment.
Over the course of model development, Gemma 2 27B was trained on a corpus totaling 13 trillion tokens from diverse sources. This dataset includes a blend of web documents for language fluency, programming code to enhance computational reasoning, and mathematical materials for improved problem-solving abilities. The model pipeline integrates automated filtering for certain categories of content, including sensitive personal data, in accordance with Google's policies.
Gemma 2 27B has been evaluated across standard benchmarks to assess its capabilities. On general language understanding and reasoning tasks, it demonstrates performance increases over the smaller Gemma 2 9B variant. For example, on the MMLU benchmark, Gemma 2 27B achieves a 5-shot top-1 accuracy of 75.2%, compared to 71.3% for the Gemma 2 9B model. Performance results are also available for a variety of tasks, including HellaSwag (86.4%), PIQA (83.2%), ARC-c (71.4%), and BIG-Bench (74.9%).
In programming and mathematics evaluation, the model reports a HumanEval pass@1 score of 51.8%, outperforming its Gemma 2 9B counterpart. For GSM8K, a mathematics-focused test, the 5-shot result is 74.0%. The model likewise undergoes assessment on safety-oriented benchmarks, such as RealToxicity, BBQ, Winogender, and Toxigen, with performance metrics aligning with established standards.
Gemma 2 27B is designed for a broad spectrum of natural language generation applications. In content creation, the model can support automated drafting of creative writing, technical documents, code, and marketing materials. Its proficiency in dialogue generation and question answering is suitable for integration into conversational agents, virtual assistants, and chatbots. Summarization capabilities enable the distillation of research literature, corporate documentation, and user-generated content.
The model also finds utility in educational and research settings, serving as a platform for natural language processing research, grammar correction tools, and automated tutoring systems. Its aptitude in code and mathematics tasks positions it as an assistive agent for programming education and computational problem-solving.
Gemma 2 27B expands the Gemma family, which comprises several related large language model variants. Notably, the Gemma 2 9B model represents a smaller-scale pre-trained and instruction-tuned alternative. Later iterations, such as those in the Gemma 3 series, introduce multimodal inputs, extended language support, and larger context windows. Closely aligned models include CodeGemma, which is specialized for code generation, PaliGemma 2 for visual data processing, and ShieldGemma 2 for safety evaluations. All Gemma models trace their research lineage to the advancements established by Google's Gemini foundation models.
Although Gemma 2 27B demonstrates notable performance across multiple benchmarks, it is subject to several well-known limitations inherent to large language models. The quality and diversity of the training corpus influence its ability to respond accurately across domains, and biases present in the data may be reflected in model outputs. As a statistical model, it can occasionally produce incorrect or outdated factual information and may encounter difficulties with ambiguous language or nuanced tasks such as sarcasm or figurative speech. Performance may also decrease when provided with insufficient context or ill-defined prompts. Additionally, the model is not equipped for common sense reasoning in the manner of a human agent, relying exclusively on observed patterns in the training data.
Access to Gemma 2 27B is governed by Google's license terms, which require users to agree to responsible use policies prior to deployment. The model documentation is released under the Creative Commons Attribution 4.0 License, while example code is subject to the Apache 2.0 License. Details of prohibited uses, as well as recommended practices for ethical deployment, are outlined within the Gemma Prohibited Use Policy and Responsible Generative AI Toolkit.