DeepSeek updates V4.1 Flash to 1017 build on Oct 17 with MIT-licensed open weights, Terminal-Bench 4.0 rises to 42.3
DeepSeek-V4.1-Flash-1017 is open-sourced on Hugging Face and ModelScope with MIT license; API price unchanged; deepseek-flash model id serves the new weights; optimized for DeepSeek Harness and other platforms.
On October 17, DeepSeek released DeepSeek-V4.1-Flash-1017 weights on Hugging Face and ModelScope under the MIT license. The model architecture, 763B parameter count, and 1M token context window remain unchanged; the update changes weights and training with self-feedback reinforcement learning. The official web, mobile app, and API now serve the new version by default.
DeepSeek said the API name deepseek-flash and pricing remain unchanged: $0.15 input / $0.60 output per 1M tokens (off-peak), $0.003 cache read; peak hours (UTC 01:00-04:00, 06:00-10:00, Monday-Friday) double the price. OpenRouter lists the new build as deepseek/deepseek-v4.1-flash-1017 at the same price. Deprecated names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to the new weights.
DeepSeek's own benchmarks show Terminal-Bench 4.0 rising to 42.3 (vs 31.2 for September), DeepSWE v1.1 to 77.0 (74.2), ProgramBench to 29.1 (20.3), HLE without tools to 42.6 (36.8), NL2Repo-Bench to 68.2 (65.4). All figures are vendor-reported; no third-party board has independently scored the 1017 build.
DeepSeek simultaneously opened its in-house search engine API, which had previously powered its own application and API search. The update is optimized for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy, and other platforms, with DeepSeek Harness 0.2.1-rc.1 or later recommended for full optimization.