5 hours ago
- DeepSeek-V4-Flash-0731 can be used with Transformers via pipeline or loading model directly, with code examples provided.
- It is also deployable using vLLM (pip install or Docker) with DSpark speculative decoding enabled via a specific flag.
- SGLang serves the model via pip or Docker, and the Docker Model Runner command is provided.
- Benchmark results show DeepSeek-V4-Flash-0731 outperforms previous versions and is competitive with proprietary models like Opus-4.8.
- The model uses a custom chat template with a dedicated encoding folder, and supports three reasoning effort levels: low, high, max.
- Local deployment recommendations include sampling parameters (temperature=1.0, top_p=0.95 for agentic tasks) and a max output length of 384K tokens.
- The model is licensed under MIT License, and citation information is provided.