DeepSeek releases V4.1-Flash-1017 with MIT-licensed open weights and stronger agent scores

Chinese AI company DeepSeek has published DeepSeek-V4.1-Flash-1017 on Hugging Face. The company describes it as a minor update to DeepSeek-V4.1-Flash that improves reasoning, coding, agentic and visual tasks while cutting task execution time and output length, so users can finish tasks faster and more cheaply.

What the model card says

• The repository and the model weights are released under the MIT licence, one of the most permissive open-source licences.
• Hugging Face lists the model size as about 763 billion parameters.
• Recommended settings include a 1-million-token context window and a maximum output of at least 256,000 tokens.
• DeepSeek-reported scores rise from 31.2 to 42.3 on Terminal-Bench 4.0, from 74.2 to 77.0 on DeepSWE v1.1 and from 36.8 to 42.6 on Humanity’s Last Exam without tools, compared with the previous V4.1-Flash.
• The release has no Jinja chat template; instead DeepSeek ships a Python reference encoder and deepseek-recipe, a set of Rust libraries with Python bindings for prompt formatting.

Context

In DeepSeek’s own table the update beats several open and closed rivals on some agent benchmarks but still trails others, such as Anthropic’s Opus 5 on Humanity’s Last Exam and Terminal-Bench. All figures are self-reported by DeepSeek and have not yet been independently verified.