DeepSeek dropped DeepSeek-V4.1-Flash-1017 on Hugging Face, a “minor update” to V4.1 Flash with big gains in reasoning, coding, agentic tasks and vision. Weights are under the MIT license.
The numbers (from DeepSeek’s model card)
- Terminal-Bench 4.0: 42.3, up from 31.2, ahead of GPT-5.6 Sol (39.9) and behind Claude Opus 5 (51.8).
- DeepSWE v1.1: 77.0, the top score in its table, above Opus 5 (74.0) and GPT-5.6 Sol (73.0).
- Automation-Bench: 60.5 vs 54.8 for the previous Flash, and above every other model in the table.
- It also cuts task time and output length, so runs finish faster and cost less.
Running it
1M-token context, max_tokens of 256K or more, temperature 1.0 and top_p 1.0 recommended. There’s no Jinja chat template; DeepSeek ships a reference Python encoder and a Rust toolkit, deepseek-recipe, that converts Messages, Chat Completions and Responses API requests. These are DeepSeek’s own benchmark numbers, so wait for independent testing.
My take
An MIT-licensed model beating frontier closed models on agentic coding is a big deal for anyone building their own agent stack.
