DeepSeek V4.1 Flash-1017

Headlines · 1

1. 7h
   DeepSeek ships V4.1 Flash-1017: faster, cheaper agentic runs plus open weights

   DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, switching its website, app and API to the new snapshot and publishing MIT-licensed weights on Hugging Face and ModelScope. The company says the update improves reasoning, coding, agentic and visual benchmarks over the previous V4.1 Flash while cutting task execution time and output length, calling it a performance-cost Pareto improvement; its own numbers put it ahead of larger models such as Claude Opus 5, GPT-5.6 Sol and Kimi K3 on Terminal-Bench 4.0 and DeepSWE v1.1. DeepSeek says reinforcement learning with self-feedback strengthened end-to-end project building from minimal prompts, including game and 3D work, and that the model is tuned for DeepSeek Harness (0.2.1-rc.1 or later recommended), Tencent CodeBuddy and WorkBuddy. It also opened its in-house search engine to developers via Search and Fetch APIs at 5 yuan per 1,000 calls, with existing tool calls in Harness, Claude Code and Codex becoming billable from November 1, 2026 (Beijing time).

   DeepSeek · Hugging Face · 🔥 95

Experience and discussion from the community

Share my DeepSeek V4.1 Flash-1017 experienceAsk about DeepSeek V4.1 Flash-1017