DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real

Summary

DeepSeek released V4.1-Flash-1017, a minor update to V4.1-Flash trained further with reinforcement learning from self-feedback for end-to-end project building. It is live on DeepSeek's website, app, and API, with MIT-licensed weights on Hugging Face (model card) and ModelScope. DeepSeek also opened Search and Fetch APIs backed by its own search engine.

Key points

• Agentic coding gains on DeepSeek's own benchmarks (1017 vs V4.1-Flash): Terminal-Bench 4.0 42.3 vs 31.2, DeepSWE v1.1 77.0 vs 74.2, ProgramBench 29.1 vs 20.3, HLE without tools 42.6 vs 36.8.
• DeepSeek claims "a Pareto improvement in performance and cost" and shorter outputs; it lists no new token prices.
• 1M-token context; recommended sampling is temperature 1.0, top_p 1.0, and max_tokens of at least 256K.
• The Hugging Face page lists 763B parameters, while DeepSeek gave V4.1-Flash 552B; the difference is not explained, and active parameters are not stated.
• Search and Fetch endpoints cost $0.75 per 1,000 calls each. Calls through server-side tools in DeepSeek Harness, Claude Code, and Codex are free until 2026-10-31 16:00 UTC.
• Tuned for DeepSeek Harness (v0.2.1-rc.1 or later), Tencent CodeBuddy, and WorkBuddy.

Why it matters

The leading open-weight lab is iterating on agentic coding with dated point releases, like closed providers do. Bundled search tools priced per call bring DeepSeek's API closer to the hosted tool offerings of closed providers.