NextLUCA
← ← Back to blog

DeepSeek-V4-Flash-0731: The Official Release Replacing the DeepSeek-V4-Flash Preview

DeepSeek has published DeepSeek-V4-Flash-0731, an official release that supersedes the earlier DeepSeek-V4-Flash preview. It shares its model structure with DeepSeek-V4-Flash-DSpark and ships with an attached speculative decoding module, benchmark results across agentic and coding tasks, and integration paths for vLLM and SGLang.

DeepSeek has published DeepSeek-V4-Flash-0731, an official release that supersedes the earlier DeepSeek-V4-Flash preview. It shares its model structure with DeepSeek-V4-Flash-DSpark and ships with an attached speculative decoding module, benchmark results across agentic and coding tasks, and integration paths for vLLM and SGLang.

Specs

  • Model structure: Same as DeepSeek-V4-Flash-DSpark
  • Reasoning effort levels: low, high, max
  • License: MIT License
  • vLLM speculative method: dspark (7 speculative tokens in the provided example)
  • SGLang speculative algorithm: DSPARK
  • Recommended max output length: 384K tokens for high and max reasoning effort
  • DeepSeek-V4-Flash-0731 is the official release superseding the DeepSeek-V4-Flash preview.
  • It uses the same model structure as DeepSeek-V4-Flash-DSpark and includes an attached speculative decoding module.
  • Benchmark scores include: 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, 70.3 on Toolathlon-Verified, 25.2 on Agents' Last Exam, 25.1 on AutomationBench Public, 68.7 on DSBench-FullStack, and 59.6 on DSBench-Hard.
  • The release does not include a Jinja-format chat template, but does provide an encoding folder with Python scripts and test cases.
  • The reasoning_effort parameter supports low, high, and max levels.
  • Speculative decoding via DSpark can be enabled in vLLM using the dspark method, with the provided example using seven speculative tokens.
  • SGLang enables DSpark through the DSPARK speculative algorithm, using target and draft weights from the same checkpoint in the provided configuration.
  • Model weights are released under the MIT License.
  • Engineers evaluating agentic coding and tool-use models against benchmarks like Terminal Bench 2.1, NL2Repo, Cybergym, and DeepSWE.
  • Teams already running vLLM or SGLang inference stacks who want to test the dspark speculative decoding integration.
  • Developers who need an MIT-licensed model and are comparing scores on DSBench-FullStack, DSBench-Hard, and AutomationBench Public before selecting a backend.
  • Anyone tracking the DeepSeek-V4-Flash line who wants to understand what changed between the preview and this official 0731 release.

The model card and weights are published on Hugging Face at deepseek-ai/DeepSeek-V4-Flash-0731 under the MIT License. To reproduce the speculative decoding setup described in the model card, configure vLLM with the dspark speculative method and seven speculative tokens as shown in the provided example, or configure SGLang with the DSPARK speculative algorithm using target and draft weights from the same checkpoint. Note that no Jinja chat template ships with the release, so you will need to supply your own prompting format, and the encoding folder with Python scripts and test cases is available for reference. If you want help evaluating this release against your existing agent or coding pipeline, NextLUCA's team tracks model releases like this one as part of its applied AI training and consulting work — book a call to talk through where it fits.

Sources