DeepSeek ships V4.1-Flash-1017 update, claiming faster, cheaper agent tasks
DeepSeek released DeepSeek-V4.1-Flash-1017, a small update to the V4.1-Flash model it launched in September. The Hugging Face model card calls it a "minor update" that "reduces task execution time and output length" without stating by how much; 36Kr reports the release keeps the original architecture, with 552B total parameters and 8B input / 16B output activation. DeepSeek's own published numbers show Terminal-Bench 4.0 rising from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and DeepSWE v1.1 task resolution from 74.2% to 77%, while AIbase cites an Arena leaderboard placing it at 1670 points, first among open models on Code Arena WebDev and 12th overall. The model is live in DeepSeek's web and mobile apps and via the existing deepseek-flash API at unchanged prices, with MIT-licensed weights on Hugging Face, and DeepSeek also opened its in-house search engine as Search and Fetch APIs at 5 yuan (about $0.75) per 1,000 successful requests.
36Kr's analysis frames the release as evidence that post-training and leaner reasoning output can lift capability and cut per-task cost without waiting for a new base architecture.
Chinese artificial intelligence startup Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. has released an updated version of its V4.1-Flash model that it says completes tasks faster.
DeepSeek-V4.1-Flash-1017 arrives a little over five weeks after the company launched the original V4.1-Flash as the smallest model in a new architecture family. Its model card on Hugging Face calls the release a “minor update” that “reduces task execution time and output length.” DeepSeek did not say by how much.
Price was the selling point in September, with off-peak application programming interface rates starting at 15 cents per million uncached input tokens. After that release, Bloomberg reported Saturday, shares of Chinese rivals MiniMax Group Inc. and Alibaba Group Holding Ltd. fell in Hong Kong.
The update resolved 77% of tasks on the DeepSWE v1.1 software engineering test in DeepSeek’s own testing, up from 74.2% for the September version. Anthropic PBC’s Claude Opus 5 sits at 74% in the same table. Seven of the 11 benchmarks DeepSeek published still go to Anthropic’s model, and the margin on Humanity’s Last Exam runs to almost 14 points.
#DeepSeek#DeepSeek-V4.1-Flash-1017#Hugging Face#Anthropic#Claude Opus 5#GPT-5.6 Sol#Tencent#CodeBuddy#WorkBuddy#Arena