模型與程式代理
DeepSeek-V4-Flash 0731 Redoes Only Post-Training, Raising Its Official Terminal-Bench 2.1 Score to 82.7
DeepSeek has moved the newly post-trained V4-Flash into public API beta without changing its architecture, 284B total parameters, or 13B active parameters. The new version natively supports the Responses API and targets coding agents and tool use, but key benchmarks rely on the unreleased DeepSeek Harness.
On July 31, DeepSeek moved the production DeepSeek-V4-Flash API into public beta. Clients still call `deepseek-v4-flash`, but the backend has been switched to V4-Flash-0731. This is not a new base model: DeepSeek explicitly states that it retains the preview version’s architecture and scale and has only undergone new post-training. It remains a Mixture-of-Experts (MoE) model with 284B total parameters and 13B active parameters per token, supporting a 1 million-token context window. The official model card provides mixed FP4/FP8 weights, with the MoE experts in FP4 and most other components in FP8.
The update focuses on agentic capabilities. DeepSeek reports scores of 82.7 on Terminal-Bench 2.1, 76.7 on CyberGym, and 70.3 on Toolathlon Verified, along with 54.2 on NL2Repo and 54.4 on DeepSWE. The API also natively accepts the Responses API format and has been tuned for Codex-style workflows, reducing protocol-conversion overhead when existing coding agents switch endpoints. DeepSeek also emphasizes that this update applies only to the Flash API; V4-Pro, the web version, and the app continue to use the previous model.
Engineering teams should not treat the reported scores as directly reproducible measures of model capability. The coding-agent evaluations used the unreleased DeepSeek Harness in minimal mode, configured with maximum reasoning effort, `top_p=0.95`, and a temperature of 1.0. DSBench-FullStack and DSBench-Hard are also internal benchmark sets. A more credible next step is to wait for the harness, complete trajectories, and cost data to be released, then compare other models using identical containers, token budgets, and retry policies. At the same time, the fact that post-training alone can substantially change tool-use performance on a fixed architecture shows that version pinning for agent deployments should record more than the model name: it should also preserve the exact snapshot, system prompt, and reasoning settings.