DeepSeek has pushed out another refinement to its lightweight model line, and the turnaround is quick even by AI industry standards. The Hangzhou-based developer released V4.1-Flash-1017 on 17 October, roughly five weeks after the original V4.1-Flash arrived in early September. The company is calling it a minor update, but the headline pitch is speed: faster task completion, shorter responses and better benchmark scores, all at the same price.
The new version is live across DeepSeek’s website, mobile app and developer platform. According to the company’s technical documentation, the gains span reasoning, coding, agentic workflows and visual tasks, with reported improvements in use cases ranging from game creation to 3D modelling. On Humanity’s Last Exam, a difficult reasoning benchmark, DeepSeek reports a score range of 36.8 to 42.6, compared with 36.8 to 39.1 for the previous release. The lower bound hasn’t moved, so the improvement looks more like a higher ceiling than a consistent lift.
Coding is where the figures get more pointed. The model reportedly scores 77.0% on the DeepSWE v1.1 coding test, up from 74.2%, and the same reports place Anthropic’s Claude Opus 5 at 74.0% on that benchmark. That comparison will grab attention, but it remains a single benchmark, and company-reported numbers rarely tell the full story of how a model behaves in real development work.
The more practical change may be the shorter outputs. API pricing is reportedly unchanged, and developers can keep using the existing deepseek-flash endpoint. Since API costs scale with generated tokens, leaner answers could translate into smaller bills without any price cut. The weights are also reportedly available under the permissive MIT licence on Hugging Face and ModelScope, keeping DeepSeek firmly in the open-weight camp. Alongside the model, DeepSeek is said to be opening its Search & Fetch APIs to outside developers for the first time, at $0.75 per 1,000 calls, with some tools free until 1 November 2026.
Pricing has been DeepSeek’s sharpest weapon. The September release was discounted so aggressively that shares of Chinese rivals including MiniMax, Z.ai and Alibaba dropped in Hong Kong trading. That pressure is now aimed squarely at expensive proprietary models from US labs.
Reliability is the weak spot. DeepSeek has faced slowdowns and service disruptions in recent months, including an outage of around five hours earlier this week that some experts linked to limited computing resources. Seen that way, shorter outputs are as much about easing strain on infrastructure as impressing benchmark charts. DeepSeek also hasn’t clarified whether app and web users are moved to the new model automatically, so checking the model label is the safest bet.
