DeepSeek-V4-Flash-0731: Flash scores 82.7, surpassing Pro

DeepSeek-V4-Flash-0731 is the official release replacing the previous preview version, a text generation model with significantly enhanced agent capabilities. It shares the same architecture as DeepSeek-V4-Flash-DSpark and includes a built-in speculative decoding module to boost inference speed. Released under the MIT license, it is available for commercial use.

Despite having far fewer active parameters, it achieves higher benchmark scores than DeepSeek-V4-Pro(Preview). It scored 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE, a level competitive with top-tier proprietary models such as GLM-5.2 and Opus-4.8. It shows particularly strong performance in code agent and tool-calling tasks.

It provides a dedicated Python script that encodes messages in an OpenAI-compatible format instead of using Jinja-based chat templates. The reasoning_effort parameter can be set to low, high, or max, allowing direct control over the inference time the model allocates before generating a response. It is compatible with major inference frameworks such as vLLM, SGLang, and Transformers, and supports 8-bit and FP8 precision options.

It is suitable for developing agents that understand complex codebases or perform automation tasks. It is useful for developers who need to maintain cost efficiency in scenarios requiring high inference accuracy. It delivers optimal performance when combined with agent frameworks like DeepSeek Harness and run with max reasoning effort.