Chinese artificial intelligence company DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, publishing the model weights on Hugging Face and ModelScope under the MIT license, which allows free commercial use. The number in the name is the release date. The model is also live on DeepSeek's website, app and API.
DeepSeek describes the model as a minor update to V4.1-Flash, released in September, with gains in reasoning, coding, agent tasks and visual benchmarks. The company says it scores higher, finishes tasks faster and writes shorter answers than the earlier version. On Terminal-Bench 4.0 it rose to 42.3 from 31.2, on DeepSWE v1.1 to 77.0 from 74.2, on ProgramBench to 29.1 from 20.3, and on Humanity's Last Exam without tools to 42.6 from 36.8.
Hugging Face lists the checkpoint at about 763 billion parameters. Technical documentation from NVIDIA for the V4.1-Flash family describes a sparse mixture-of-experts backbone of roughly 552 billion parameters, with 384 routed experts and six selected for each token, alongside about 196 billion parameters in a memory component DeepSeek calls Engram. The model card recommends a context window of 1 million tokens and a maximum output of at least 256,000 tokens.
Alongside the model, DeepSeek opened its Search and Fetch APIs to all developers. The tools, which let the model look up and read pages on the live web, run on DeepSeek's in-house search engine and cost $0.75 per 1,000 calls for each endpoint. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts on October 31 at 16:00 UTC. DeepSeek recommends Harness version 0.2.1-rc.1 or later for the best results.
The benchmark scores are DeepSeek's own and had not been independently reproduced at release. Community developers have already published compressed versions of the weights on Hugging Face, also under the MIT license.
