DeepSeek releases V4.1 Flash 1017 with gains in coding and agent tests

Expand

Open original image ↗

DeepSeek has released V4.1 Flash 1017, an update to its Flash model that improves coding, reasoning and visual results while, the company says, taking less time and producing shorter outputs. It is available now on DeepSeek’s website, app and API, with downloadable weights on Hugging Face and ModelScope. 1

DeepSeek calls it a minor update. The comparison with the previous version is nevertheless encouraging: every benchmark listed in its model card improves. The biggest jumps are on Terminal-Bench 4.0, up from 31.2 to 42.3, and ProgramBench, which rises from 20.3 to 29.1. Those are gains of 11.1 and 8.8 points respectively. 2

More tasks completed, at a lower reported cost

On DeepSWE v1.1, Flash 1017 reaches 77.0, compared with 74.2 for the earlier Flash release. DeepSeek’s cost chart puts the new model’s highest-effort result at roughly $0.20 per task, against roughly $0.50 for the previous version’s best result. That is a useful combination for coding agents, where retries and long sequences of tool calls can make the cost of finishing a task more revealing than the price of a single response.

The gains extend beyond software engineering. Automation-Bench moves from 54.8 to 60.5, Agents’ Last Exam from 31.8 to 35.4, and Apex Bench v1.1 from 64.6 to 72.1. HLE without tools rises from 36.8 to 42.6; its separately reported text-only subset increases from 39.1 to 44.0.

Expand

Open original image ↗

Flash 1017 leads the listed models on DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench. Claude Opus 5 remains ahead on HLE, Terminal-Bench, ProgramBench and NL2Repo-Bench. The visual scores also improve over the old Flash, including 92.5 on BabyVision with tools, though Opus scores 94.1. This is DeepSeek’s comparison; Lumina has not independently reproduced the results.

From a prompt to a complete project

DeepSeek says it used reinforcement learning with self-feedback to improve project development from start to finish. Its demonstrations include a 3D Forbidden City, a browser racing game, a Portal 2 recreation in Godot and a scene following civilisation from prehistoric times to the modern era. They give a sense of the work DeepSeek is targeting, although the selected demos do not establish how reliably the model completes unfamiliar projects.

The release is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, among other tools. DeepSeek recommends Harness v0.2.1-rc.1 or later. The model card lists a one-million-token context window, and both the repository and weights carry the MIT licence. 1 2

Search and Fetch are now available separately

Alongside the model, DeepSeek has opened two web APIs. Search returns results from its own search engine; Fetch retrieves a page as Markdown. Each costs $0.75 per 1,000 successful requests. Failed requests, empty search results and timeouts are not billed. 3

There is a temporary exemption for server-side Search and Fetch calls made through DeepSeek Harness, Claude Code and Codex by DeepSeek API users. Those calls remain free until billing begins on 31 October at 16:00 UTC. The exemption applies to those tool calls; the separate APIs have the published per-request price. 1 3