DeepSeek-V4.1-Flash-1017: better coding, shorter outputs

What was released?

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17. It is a minor update to V4.1-Flash, which launched on September 10 (Hugging Face). The company says the new version scores higher on benchmarks, completes tasks faster and produces shorter outputs. DeepSeek calls this a Pareto improvement in both performance and cost (DeepSeek).

The model is live on DeepSeek's website, app and API. The weights are available under the MIT License on Hugging Face and ModelScope (ModelScope). The model card lists a 1-million-token context window and recommends max_tokens of at least 256K.

Benchmarks versus the previous V4.1-Flash

In the table DeepSeek published, the 1017 version improves on its predecessor on all 11 benchmarks:

• HLE without tools: 42.6 vs. 36.8 (44.0 vs. 39.1 on the text-only subset);
• Terminal-Bench 4.0: 42.3 vs. 31.2;
• DeepSWE v1.1: 77.0 vs. 74.2;
• ProgramBench: 29.1 vs. 20.3;
• NL2Repo-Bench: 68.2 vs. 65.4;
• Automation-Bench v1.0.6: 60.5 vs. 54.8;
• Agents' Last Exam: 35.4 vs. 31.8;
• Apex Bench v1.1: 72.1 vs. 64.6;
• visual benchmarks with tools: Chartography 83.0 (78.9), BabyVision 92.5 (89.6), ZeroBench-main 51.0 (49.0).

Source: DeepSeek. Benchmark results for DeepSeek-V4.1-Flash-1017 compared with other models, as measured by the vendor.

The biggest gains are on Terminal-Bench 4.0 and ProgramBench, up 11.1 and 8.8 points. Terminal-Bench was run with the Claude Code harness and DeepSWE with Mini-SWE-V2. The ZeroBench-main score is Pass@5.

The table is not all good news. Claude Opus 5 leads on six benchmarks, including HLE at 56.3 and Terminal-Bench at 51.8. GPT 5.6-Sol leads ZeroBench-main at 53.0. DeepSeek's model comes out on top on DeepSWE, Automation-Bench, Agents' Last Exam and Apex Bench. These are the vendor's own measurements, with no independent verification yet.

Faster and cheaper per task

The headline figure in the announcement plots DeepSWE v1.1 Pass@1 against average cost per task. At max effort, the 1017 version reaches 77.0%. Its curve sits to the left of and above the previous V4.1-Flash curve at every effort level, meaning better results for less money.

Source: DeepSeek. DeepSWE v1.1 Pass@1 vs. average cost per task, using the mini-swe-v2 harness.

DeepSeek gave no specific figures for speed or output length, only that tasks finish faster and outputs are shorter. The announcement does not mention any change to token prices. For agent workloads, shorter outputs still translate directly into a smaller bill.

Agents, harnesses and demos

The model is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later for the best experience. The company says it used reinforcement learning with self-feedback to improve building complete projects from scratch.

It showed four demos: a 3D model of the Forbidden City, a browser-based racing game, a Portal 2 recreation built in the Godot engine and a 3D scene tracing history from prehistoric times to the present (Valley of Ages).

Search and Fetch APIs open to everyone

Alongside the model, DeepSeek opened its Search and Fetch APIs, powered by its in-house search engine, to all users. Each endpoint costs $0.75 per 1,000 calls. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are currently free for DeepSeek API users. Billing starts on October 31, 2026, at 16:00 UTC (DeepSeek).

Who should take a look?

Teams already running V4.1-Flash with a coding agent get an easy swap: same model family, higher scores, shorter outputs. The open weights also make it possible to test on your own infrastructure. As always, rerun the benchmarks on your own tasks before switching anything over.