DeepSeek ships V4.1-Flash-1017, a faster update with shorter outputs

DeepSeek has released an update to its V4.1-Flash model, pitching the new version as faster and more cost-effective. According to Bloomberg, the Hangzhou-based company unveiled V4.1-Flash-1017 late on Saturday, 17 October, in a series of posts on X.

DeepSeek said the update has “stronger benchmark performance, faster task completion, and shorter outputs than the previous version.” The company said the model is available on its website, in its mobile app and on its developer platform.

The update follows quickly on the heels of its predecessor. V4.1-Flash, the slimmed-down system it replaces, arrived in early September. DeepSeek describes the new release as a minor update to that model. Bloomberg’s report, republished by The Straits Times, says it improves capabilities in uses ranging from game creation to 3D modelling. A technical report on Hugging Face claims “substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” On Humanity’s Last Exam, the revised model scored 36.8-42.6, against 36.8-39.1 for V4.1-Flash.

Coding is where the numbers get most specific. Crypto Briefing reports that the model scored 77.0% on the DeepSWE v1.1 coding test, up from 74.2% for the previous version, and that the same benchmarks put Anthropic’s Claude Opus 5 at 74.0%.

The outlet also says API pricing is unchanged and developers can access the model through the existing deepseek-flash endpoint, so shorter outputs should lower bills for anyone paying per generated token. It adds that the weights are open-sourced under the MIT licence on Hugging Face and ModelScope, and that DeepSeek is opening its Search & Fetch APIs to outside developers for the first time. The same report says those APIs cost $0.75 per 1,000 calls, with some tools free until 1 November 2026.

The release lands in a crowded field. Bloomberg says DeepSeek’s launch pace underscores the contest between China’s open-weight models and the costlier proprietary models from US rivals. The September model was offered at such a steep discount that shares of rivals including MiniMax Group, Z.ai and Alibaba fell in Hong Kong.

A copy of the report carried by Red Hot Singapore adds that DeepSeek has seen slowdowns and service disruptions in recent months, putting pressure on its services. The Bloomberg report also describes an outage earlier that week that lasted about five hours, with some experts attributing it to insufficient computing resources. Speed is therefore as much a capacity question as a benchmark one.

For people who use DeepSeek’s app or website, DeepSeek’s announcement doesn’t spell out whether the new model replaces the old one by default or how it rolls out by country. We’d check the model label in the app before assuming you’re on the new version.