ai/deepseek-v4-flash-safetensors repository overview

⁠DeepSeek-V4-Flash

DeepSeek-V4-Flash is a highly efficient Mixture-of-Experts (MoE) language model with 284B total parameters, of which 13B are activated per token. It is part of the DeepSeek-V4 series, designed to deliver exceptional performance with significantly improved efficiency for long-context tasks. The model supports an unprecedented context length of one million tokens, making it ideal for processing large documents, extensive codebases, and complex multi-turn conversations.

Built with innovative architectural advances including Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), DeepSeek-V4-Flash achieves world-leading long-context efficiency while drastically reducing computational and memory costs. Pre-trained on over 32 trillion diverse, high-quality tokens and refined through a comprehensive two-stage post-training pipeline, the model delivers strong reasoning capabilities that closely approach its larger sibling DeepSeek-V4-Pro, particularly when given a larger thinking budget. Its smaller parameter footprint translates to faster response times and highly cost-effective deployment, making it an excellent choice for production environments requiring both performance and efficiency.

DeepSeek-V4-Flash supports dual reasoning modes: a fast non-thinking mode for routine tasks and a thinking mode for complex problem-solving that rivals the Pro version's performance. The model excels across diverse domains including coding, mathematics, agentic workflows, and knowledge-intensive tasks, establishing itself as a top-tier open-source model in its parameter class.

⁠Characteristics

⁠Using this model with Docker Model Runner

docker model run deepseek-v4-flash-safetensors

For more information, check out the Docker Model Runner docs⁠.

⁠Benchmarks

⁠Base Model Performance

⁠Instruct Model Performance (Across Reasoning Modes)

⁠Key Features

Hybrid Attention Architecture: DeepSeek-V4-Flash employs a novel hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). In the 1M-token context setting, this design requires only 27% of single-token inference FLOPs and 10% of KV cache compared to DeepSeek-V3.2.

Manifold-Constrained Hyper-Connections (mHC): The model incorporates mHC to strengthen conventional residual connections, enhancing stability of signal propagation across layers while preserving model expressivity.

Muon Optimizer: Training leverages the Muon optimizer for faster convergence and greater training stability.

Reasoning Modes: The model supports three reasoning effort modes:

• Non-think: Fast, intuitive responses for routine daily tasks and low-risk decisions
• Think High: Conscious logical analysis, slower but more accurate for complex problem-solving
• Think Max: Push reasoning to its fullest extent for exploring the boundary of model capabilities

⁠Considerations

• For local deployment, the model requires at least 384K token context window for Think Max reasoning mode
• Recommended sampling parameters: temperature = 1.0, top_p = 1.0
• The model uses FP4 precision for MoE expert parameters and FP8 for most other parameters, optimizing for both performance and efficiency
• While DeepSeek-V4-Flash delivers exceptional reasoning performance approaching the Pro version, its smaller parameter scale naturally places it slightly behind on pure knowledge tasks and the most complex agentic workflows
• Specialized encoding scripts are required for message formatting (provided in the repository)

⁠Links

• https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash⁠
• https://api-docs.deepseek.com/news/news260424⁠
• Technical Report⁠
• DeepSeek Chat⁠
• GitHub Repository⁠

⁠Generated by

This model card was automatically generated using cagent-action⁠. Want to learn more about Docker Model Runner? Check out the project repository: https://github.com/docker/model-runner⁠.