DeepSeek launched the DeepSeek-V4.1-Flash-1017 model and published its weights on ModelScope and Hugging Face. The company claims the new version delivers superior benchmark performance compared to its predecessor.
Developers used reinforcement learning with self-feedback (RLSF) during post-training to enhance the model's capacity to build complete projects from scratch. Tuning for the V4.1-Flash-1017 included DeepSeek Harness, WorkBuddy, and Tencent CodeBuddy.
DeepSeek asserts that V4.1-Flash-1017 beats Claude Opus 5, GPT-5.6 Sol, Kimi K3, and GLM 5.3 across various AI agent and coding benchmarks. The model achieved a peak score of 77.0% on the DeepSWE benchmark at a cost of just under $0.20 per task.
At a low setting, the model scored above 72% for approximately $0.10 per task, whereas the previous version reached 74.2% at roughly $0.50 per run. DeepSeek is also making its internal web search technology available to external developers via Search and Fetch APIs priced at $0.75 per 1,000 calls.
API users can access these tools for free through server-side integrations until 4:00 PM UTC on 31 October 2026. Reports suggest the Chinese government may impose tighter controls on publishing open-weight AI models, with a proposal potentially restricting the 10 most capable systems.
