DeepSeek makes its cheapest model more powerful: DeepSeek-V4-Flash

DeepSeek-V4-Flash is a lighter, more cost-effective model in the DeepSeek-V4 series optimized for faster inference. Compared to Pro and Ultra versions, it has fewer activated parameters (304B) but achieves competitive benchmark results. The model includes a speculative decoding module (DSpark) to accelerate text generation.

DeepSeek-V4-Flash-0731 shows significant improvements in agentic benchmarks – for example, Terminal Bench 2.1 improved from 61.8 to 82.7, and DeepSWE from 7.3 to 54.4. The model also supports three reasoning_effort levels (low, high, max) to control the depth of deliberation.

The model can be deployed using vLLM or SGLang with speculative decoding support enabled. For local deployment, scripts are provided in the inference folder. Recommended settings are temperature=1.0 with top_p=0.95 for agentic scenarios and top_p=1.0 otherwise, with maximum output length of 384K tokens for higher reasoning effort levels.

The model is licensed under the MIT License, allowing free commercial and non-commercial use. Weights and code are publicly available on Hugging Face, and the community has already created 75 quantized versions and 11 fine-tuned models.