DeepSeek releases V4.1-Flash-1017 under MIT licence

DeepSeek has released V4.1-Flash-1017, an update to its V4.1 Flash model, on Hugging Face, with the repository and weights under the MIT licence.

DeepSeek said in the model card that the release is a minor update with "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks". It said the model also cuts task execution time and output length, which lowers cost.

The context window is 1M tokens. DeepSeek recommends a max_tokens setting of 256K or more, with temperature 1.0 and top_p 1.0.

The card compares Flash-1017 with the original Flash. On Humanity's Last Exam without tools (HLE), it scores 42.6 against 36.8. On Terminal-Bench 4.0 it scores 42.3 against 31.2, and on ProgramBench 29.1 against 20.3.

On DeepSWE v1.1 the scores are 77.0 against 74.2. On Apex Bench v1.1 they are 72.1 against 64.6, and on BabyVision 92.5 against 89.6.

The card also lists results for K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Opus 5 scores higher than Flash-1017 on most of the listed benchmarks.

The release includes no Jinja chat template. DeepSeek provides a Python reference implementation, encoding.py, and a Rust toolkit, deepseek-recipe, for production use.

The model card gives no release date. The 17 October date comes from a news roundup on aiweekly.co, which also lists 552B total parameters and new Search and Fetch APIs at $0.75 per 1,000 successful requests. The model page shows 763B parameters and lists no API pricing. DeepSeek has not confirmed the parameter count or the API pricing.