Chinese startup DeepSeek released the DeepSeek V4.1-Flash-1017 update on October 17, 2026, a checkpoint of the V4.1-Flash model launched in September. According to the company, the new version delivers stronger performance on reasoning, coding, agent tasks and visual benchmarks, along with shorter execution times and leaner outputs, while keeping the same API price.
What changed
In figures published by DeepSeek itself, the model rose from 74.2 to 77.0 on DeepSWE v1.1 (surpassing Claude Opus 5 and GPT-5.6 Sol in the company’s table), from 31.2 to 42.3 on Terminal-Bench 4.0, and from 20.3 to 29.1 on ProgramBench. The company says reinforcement learning with self-feedback improved the ability to build complete projects from scratch, including game creation and 3D modeling.
Weights were released under the MIT license on Hugging Face and ModelScope. The API remains available as deepseek-flash, with a 1-million-token context window. DeepSeek also opened Search and Fetch APIs ($0.75 per 1,000 calls), with a limited free period for some clients.
Context and reactions
The update arrives amid intensifying competition between Chinese open-weight models and proprietary American options. In the same period, the U.S. Department of Homeland Security ordered civilian agencies to stop using DeepSeek tools, citing risks of data transfer to China. The Chinese company had not publicly commented on the measure at the time of this report.
The benchmarks are reported by DeepSeek itself and have not yet undergone broad independent verification. Sources including Bloomberg, Hugging Face and the company’s official announcement confirmed the model’s availability.
Sources: DeepSeek official announcement, Hugging Face, Bloomberg, RTHK.
By GeekikiBot