DeepSeek upgrades V4.1 Flash to DeepSeek-V4.1-Flash-1017

DeepSeek says it has upgraded its V4.1 Flash model to DeepSeek-V4.1-Flash-1017. The official web and app models now use it. The API is unchanged: it is still called as deepseek-flash, and DeepSeek says pricing stays the same.

DeepSeek's claims (from its own update note, not independently verified): the new model is a "performance–cost Pareto improvement" over the previous V4.1, with shorter task runtime and output length. It reports HLE (no tools) 42.6, Terminal-Bench 4.0 42.3 (run with the Claude Code framework), DeepSWE v1.1 77.0, ProgramBench 29.1, NL2Repo-Bench 68.2 and BabyVision (with tools) 92.5. DeepSeek says it tuned the model for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends DeepSeek Harness 0.2.1-rc.1 or later.

Weights: a deepseek-ai/DeepSeek-V4.1-Flash-1017 repository (pipeline image-text-to-text) appears in the Hugging Face catalog.

Third-party figures: Artificial Analysis, a benchmark site, lists the model as open weights under the MIT license, with 552B total and 16B active parameters and a 1M-token context window. It gives the model an Intelligence Index of 45, ranked #2 of 121 in its comparison set, and 230 output tokens per second. It also lists $0.30 per 1M input tokens and $1.20 per 1M output tokens, and calls the model very verbose. These are the site's measurements, not DeepSeek's.