DeepSeek releases V4.1-Flash-1017, an update it says is faster and produces shorter outputs

DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated version of its V4.1-Flash model. In its release announcement, the company said the new version scores higher on benchmarks, finishes tasks more quickly and generates shorter responses than its predecessor, calling the result "a Pareto improvement in performance and cost". According to DeepSeek, the model is available on its website, in its app and through its API, and the weights have been published on Hugging Face and ModelScope.

The Economic Times, carrying a wire report, said DeepSeek announced the update on Saturday afternoon in a series of posts on X and described it as a minor revision of the earlier V4.1-Flash. The report cited the company's technical report on Hugging Face as claiming "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks". According to the report, DeepSeek also posted scores on Humanity's Last Exam, a test of thousands of questions spanning math, the sciences and the humanities, giving a range of 36.8 to 42.6 for the new model against 36.8 to 39.1 for V4.1-Flash. Neither source explains what the ranges represent, and the figures have not been independently verified.

DeepSeek said the model was trained with reinforcement learning using self-feedback to get better at building full projects end to end, and it showcased examples including a 3D model of the Forbidden City, a browser-based racing game and a Portal 2 recreation made with the Godot engine. The company said the model has been optimized for tools including DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and it recommends DeepSeek Harness v0.2.1-rc.1 or newer. It also published a chart plotting DeepSWE v1.1 Pass@1 against cost per task, measured with the mini-swe-v2 harness.

Alongside the model, DeepSeek opened its Search and Fetch APIs to all users. It said they run on its own search engine, which it says handles hundreds of millions of requests a day, and priced each endpoint at $0.75 per 1,000 calls. According to the company, DeepSeek API users can call these tools for free through server-side tools in DeepSeek Harness, Claude Code and Codex until billing begins on Oct 31, 2026, at 16:00 UTC.

The report places the release in a wider contest between Chinese open-weight models and higher-cost proprietary US systems from companies such as OpenAI and Anthropic, as well as domestic rivals including Moonshot Kimi and Z.ai. It noted that the steep discount at which V4.1-Flash was offered after its early-September launch was followed by sharp drops in the shares, listed in Hong Kong, of competitors such as MiniMax, Z.ai and Alibaba. The report also said DeepSeek had a server outage of about five hours earlier in the week, which some experts linked to insufficient computing capacity, and that slowdowns and disruptions have become more frequent in recent months as developer demand has grown.