AINEWS
The Chinese AI lab DeepSeek fine tuned the same size model further on agent tasks, posting an 82.7 score on Terminal Bench 2.1, a public agent coding benchmark; the disclosure gap is now the load bearing part.
DeepSeek released an updated version of its V4-Flash language model on July 31, the company said, and the change is narrower than the benchmark headlines suggest: same architecture, same parameter count, fresh post-training. The model is now called DeepSeek-V4-Flash-0731, and it is live on the company's API as the official release of a model that had previously been in preview.
"Post-training," in the LLM context, means fine-tuning the same model further on agent and tool-use tasks, without changing its size or architecture. That is the lever DeepSeek pulled. The version bump is built for "agent" work, which here means multi-step coding tasks where the model reads a repository, runs shell commands, and fixes its own errors rather than autocompleting a snippet.
On the agent coding benchmark Terminal Bench 2.1, DeepSeek reports a score of 82.7 for the new model. The public Terminal-Bench 2.1 leaderboard shows the top entry at 83.8%, with an uncertainty band of plus or minus 1.2%, as of June 7, 2026. DeepSeek's reported number would slot just behind the leader at standard configuration.
The deeper move is on internal coding-agent tests that DeepSeek has not put on any third-party leaderboard. The company reports 68.7 on DSBench-FullStack and 59.6 on DSBench-Hard, alongside 54.4 on DeepSWE (a software-engineering benchmark from Princeton's PLAI group), 76.7 on Cybergym, 54.2 on NL2Repo, 70.3 on Toolathlon verified, 25.2 on Agent Last Exam, and 25.1 on Automation Bench. DSBench-FullStack and DSBench-Hard are DeepSeek's own test sets, not third-party reproducible, and the company has not disclosed the post-training data, reward methods, or training procedure that produced the jump.
That disclosure gap is the actual story. The Chinese trade publication Leiphone noted the same point: DeepSeek got materially better agent numbers from a model whose weights did not change, and the company has not told anyone what it did to get there. The result is a benchmark lift whose magnitude is bounded by what was measured and how.
The test conditions diverge from default deployment in ways that bias the numbers upward. Public Code Agent runs on the new model used DeepSeek's own Harness, a thin scaffolding layer the company says it will release soon, in its minimal mode at the max effort level, with top_p at 0.95 and temperature at 1.0. Those sampling settings produce more diverse, more exploratory outputs than the lower-temperature defaults most users see in chat products. The 82.7 Terminal Bench 2.1 figure is a directional data point, not an independent leaderboard claim.
The interface around the model also changed. V4-Flash now natively supports the Responses API, a unified interface for tool calls, reasoning state, and model replies designed for agent workflows. The company says the model is adapted to plug into Codex. The more interesting question is what those interface changes imply about the post-training pipeline behind them.
Three questions the next release will answer. First, whether DeepSeek publishes enough of its post-training recipe to let an outside lab reproduce the agent lift on a comparable model. Second, whether the Terminal-Bench 2.1 number holds up at default sampling settings and through the unreleased Harness, when other labs can run it. Third, whether the same post-training recipe generalizes to V4-Pro, which the company has not yet re-trained. DeepSeek's update log currently lists only the V4-Flash API endpoint as refreshed; the V4-Pro model, the consumer app, and the web product are unchanged.
The post-training-only move is the most useful artifact of the release, because it is the part a developer can actually test. If a same-size model with no architecture change can move from preview-level to near-leaderboard on multi-step coding tasks, then post-training is doing more of the agent-capability work in 2026 than the parameter-scaling narrative suggests. DeepSeek has not shown its work, and the next release of the Harness matters more than the next round of press-release numbers.