DeepSeek VL2 Tiny is a Mixture-of-Experts (MoE) Vision-Language Model that represents the most compact variant in the DeepSeek-VL2 model family. With 1.0B activated parameters, it offers an efficient solution for multimodal understanding tasks while maintaining competitive performance. The model is part of a family that includes DeepSeek VL2-Small (2.8B activated parameters) and the full DeepSeek-VL2 model (4.5B activated parameters), as detailed in the research paper.
Architecture and Technical Specifications
The model's architecture combines a SigLIP-SO400M-384 vision encoder with a DeepSeekMoE language model, utilizing a Multi-head Latent Attention (MLA) mechanism for efficient inference. A notable feature is its dynamic tiling vision encoding strategy, which enables processing of high-resolution images with varying aspect ratios by dividing them into manageable tiles. For multiple image processing, the model employs different strategies: up to two images use dynamic tiling, while three or more images are padded to 384x384 without tiling. The model supports a sequence length of 4096, making it suitable for complex multimodal tasks.
Training and Capabilities
DeepSeek VL2 Tiny was trained on a diverse vision-language dataset through multiple stages, including VL alignment, VL pretraining, and supervised fine-tuning. The training data encompasses various sources such as ShareGPT4V, WIT, WikiHow, OBELICS, and Wanjuan, along with specialized datasets for image captioning, OCR, visual question answering, and visual grounding. This comprehensive training approach enables the model to excel in:
- Visual question answering
- OCR and text recognition
- Document understanding
- Table and chart interpretation
- Visual grounding
- Multi-image conversation
- Visual storytelling
Performance and Practical Usage
The model demonstrates competitive performance compared to other open-source models of similar size, particularly in OCR-related tasks and visual grounding. For optimal generation quality, it's recommended to use a temperature (T) ≤ 0.7 during sampling. The model's implementation requires Python 3.8 or higher and utilizes the transformers library, as detailed in the GitHub repository.
When comparing the variants within the DeepSeek-VL2 family, performance generally improves with model size, though DeepSeek VL2 Tiny maintains competitive capabilities while requiring significantly fewer computational resources than its larger siblings. This makes it an attractive option for applications where resource efficiency is a priority.
Licensing and Availability
The model was released on December 13, 2024, with additional features including a Gradio demo and incremental prefilling support added on December 25, 2024. The code is available under the MIT License, while the model itself is governed by the DeepSeek Model License. Commercial use is supported under these licensing terms.
Reference Links
