DeepSeek Releases V4.1-Flash-1017 Update With MIT Licence
A minor version bump delivers big benchmark jumps, DeepSeek says
The facts
- DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, released with MIT-licensed weights on Hugging Face
- DeepSeek reports gains on nearly every listed benchmark, including Terminal-Bench 4.0 (42.3 vs 31.2) and Apex Bench v1.1 (72.1 vs 64.6)
- The model supports a 1M-token context window and a recommended minimum output of 256K tokens
- No Jinja chat template is included; DeepSeek instead ships a Python reference encoder and a new Rust/Python toolkit called deepseek-recipe
DeepSeek has released DeepSeek-V4.1-Flash-1017, a minor update to its V4.1-Flash model. The weights are published on Hugging Face under an MIT licence. The company says the update brings substantial gains across reasoning, coding, agentic and visual benchmarks. It also cuts task execution time and output length. DeepSeek describes this as pushing the performance-cost trade-off further for developers running the model through agent harnesses.
The benchmark table DeepSeek published shows the new release beating its predecessor across the board. Terminal-Bench 4.0 rises to 42.3 from 31.2. Apex Bench v1.1 climbs to 72.1 from 64.6. Automation-Bench v1.0.6 improves to 60.5 from 54.8. DeepSeek also compares the model against rivals it calls K3, GLM-5.3, GPT-5.6 Sol and Opus 5. On Terminal-Bench 4.0 the new DeepSeek model outscores all of those except Opus 5, which leads at 51.8. On Automation-Bench v1.0.6 and Agents' Last Exam it tops every model listed, including Opus 5. Opus 5 still leads comfortably on the Humanity's Last Exam text-only subset, at 56.3 against DeepSeek's 44.0.
The model ships with a one-million-token context window. It carries a recommended minimum output length of 256,000 tokens, with temperature and top-p both set to 1.0. DeepSeek is not shipping a standard Jinja chat template this time. Instead it provides a Python reference implementation of its prompt encoding. It also offers a new toolkit called deepseek-recipe, built as Rust libraries with Python bindings, to convert Messages, Chat Completions and Responses API requests into model-ready prompts. Model inference, tool execution and HTTP transport are left for developers to implement themselves.
Why it matters
The MIT licence and benchmark gains over named rivals like K3 and GLM-5.3 put pressure on other open-weight labs to match Flash-1017's cost-performance claims. But dropping the standard chat template shifts integration work onto developers. They must now adopt DeepSeek's own tooling or write their own encoder.
Written from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.