DeepSeek-V4.1-Flash-1017 was officially released on October 17, 2026, with DeepSeek rolling the version out the same day through its official WeChat account, its social accounts and the Hugging Face model page. The website, the app and the API have all switched to the new model; the API name stays deepseek-flash and pricing is unchanged from the previous version. The weights are published on Hugging Face and ModelScope under the MIT license. DeepSeek calls it a minor update, yet in the official benchmark table nearly all of the gains land on coding and agentic tasks.
The steepest gains are in coding and agent benchmarks
| Benchmark | V4.1-Flash-1017 | Previous V4.1 Flash |
|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 |
| Terminal-Bench 4.0 | 42.3 | 31.2 |
| ProgramBench | 29.1 | 20.3 |
| HLE (no tools) | 42.6 | 36.8 |
| NL2Repo-Bench | 68.2 | 65.4 |
First, the methodology: DeepSWE was measured on the Mini-SWE-V2 harness and Terminal-Bench 4.0 on the Claude Code harness, and all figures are DeepSeek-reported, with no independent replication yet. In the same table, the 1017 build's DeepSWE score sits above GPT-5.6 Sol at 73.0 and Claude Opus 5 at 74.0, and its Terminal-Bench 4.0 result beats K3 at 12.6 and GLM-5.3 at 37.9. For a lightweight Flash-tier model, that places it inside flagship score territory.
It beats Opus 5 on only some of the table
Switch the lens to general reasoning and the picture changes. On HLE without tools, Opus 5 scores 56.3 against 42.6; on Terminal-Bench 4.0, Opus 5 scores 51.8 against 42.3; on NL2Repo, Opus 5 also leads, 75.3 to 68.2. The comparison anchor is Opus 5, not the newer Opus 5.5. The accurate summary is that 1017 reaches flagship-level scores on software-engineering benchmarks while a clear gap remains on general reasoning. DeepSeek's own label for the release is, accordingly, a minor update.
The real work went into tuning for agent harnesses
Training for this version used reinforcement learning with self-feedback, and the official demos concentrate on one kind of task: give the model a single-sentence request and let it plan, write code and ship a complete project on its own, including building 3D scenes and small games from scratch. The release notes also name DeepSeek Harness, Tencent CodeBuddy and WorkBuddy as specifically optimized targets, and recommend upgrading Harness to v0.2.1-rc.1 or later. With faster task completion and shorter outputs, DeepSeek says the new version pushes the performance-cost Pareto frontier across several harnesses: the same work finishes sooner, with fewer output tokens, so the real cost per task drops.
For users, the impact splits in two. Teams calling the API directly change nothing: deepseek-flash now points at the new model, pricing is the same, and the context window remains 1 million tokens. Teams self-hosting the weights should note that this release ships no Jinja-format chat template; the repository's encoding folder provides a reference implementation, and a Rust toolkit called deepseek-recipe handles prompt encoding, so that layer needs aligning during migration.
When the previous version launched in September, OpenCode made DeepSeek V4.1 Flash free to use, letting many developers try it at no cost. The 1017 build keeps the same open arrangement: the weights, the price and the API name are unchanged. Only the scorecard moved.
