DeepSeek released an updated artificial intelligence model late Saturday Oct. 17, rolling out DeepSeek-V4.1-Flash-1017 across the company's website, mobile app and developer platform. The Chinese company says the new version is faster and leaner than the model it replaces, and it described the change as a "Pareto improvement" in both performance and cost — a term indicating the upgrade gained on one front without giving up ground on the other.

DeepSeek announced the update in posts on X, citing stronger benchmark performance, faster task completion and shorter outputs than the previous version delivered by V4.1-Flash. The release was also highlighted in an Oct. 18 post on X by the account Coin Bureau.

🚨JUST IN: China's DeepSeek releases V4.1-Flash-1017, claiming FASTER performance and improved reasoning, coding, and AI agent capabilities. The open-weight model reportedly outperforms its predecessor across nearly all benchmarks while completing tasks faster and using fewer… pic.twitter.com/G0gdBa1FiV

— Coin Bureau (@coinbureau) October 18, 2026

What the Update Includes

The company described the release as a minor update to V4.1-Flash. According to DeepSeek, the model improves reasoning, coding, agentic tasks and visual benchmarks, and its technical report on Hugging Face also points to better results in game creation and 3D modeling.

The model uses reinforcement learning with self-feedback, an approach DeepSeek says helps it build projects from start to finish. It is optimized for tools including DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.

DeepSeek also opened its Search and Fetch APIs to all users. The services run on the company's own search engine, and each Search or Fetch API endpoint costs $0.75 per 1,000 calls. Search and web-retrieval tools of this kind are commonly used to ground model answers in current information.

Benchmark Scores

DeepSeek published a set of benchmark comparisons alongside the release. The model scored 77.0 on the DeepSWE v1.1 benchmark, up from 74.2 for the earlier version. On Automation-Bench V1.0.6, it scored 60.5, compared with 54.8 previously.

The company also shared results for Humanity's Last Exam, a test that covers thousands of questions in math, science and the humanities. The revised model scored 36.8 to 42.6 on that test, while V4.1-Flash scored 36.8 to 39.1.

The gains are incremental, in line with the company's own framing of the release as a minor update rather than a new model generation.

Competitive Landscape

The release adds to DeepSeek's challenge to US rivals Anthropic and OpenAI, as well as to Chinese competitors such as Moonshot Kimi and Z.ai. DeepSeek's open-weight models cost far less than proprietary US models, although the US models remain top performers. Open-weight releases publish a model's parameters so developers can download and run them directly, a distribution model that contrasts with proprietary systems offered only through paid services.

The earlier V4.1-Flash launched in early September at a steep discount. Shares of MiniMax Group, Z.ai and Alibaba Group fell on the Hong Kong stock exchange afterward.

The update follows a difficult stretch for the company. Earlier in the week, DeepSeek experienced an outage of about five hours. Some experts blamed a lack of computing resources, and service slowdowns have increased in recent months. Reliability under demand is likely to remain a focus for users as the new rollout expands across the company's platforms.

The company began rolling out the update late Saturday, and V4.1-Flash-1017 is available now on the web, in the app and through the API.