DeepSeek releases V4.1-Flash-1017 open weights under MIT licence, claims agent gains
Representative image: doki7 / Pixabay
DeepSeek put V4.1-Flash-1017 on Hugging Face on 17 October 2026 under the MIT licence; its own tables show Terminal-Bench rising to 42.3 from 31.2.
What Happened
Chinese AI company DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated version of its V4.1-Flash large language model, with the model weights published on Hugging Face under the permissive MIT licence on Saturday, 17 October 2026. The "1017" in the name marks the release date. According to DeepSeek's own comparison tables, the update scores 42.3 on the Terminal-Bench 4.0 test of agent skills, up from 31.2 for the September version of V4.1-Flash, and also improves on two other agent benchmarks. Alongside the weights, the company published a price of 0.75 dollars per 1,000 successful requests for its Search and Fetch APIs, according to industry trackers.
Key Facts
- Model: DeepSeek-V4.1-Flash-1017, an October snapshot of V4.1-Flash, which itself was released on 10 September 2026.
- Licence: the repository and weights are under the MIT licence, which allows commercial use and modification.
- DeepSeek's figures: Terminal-Bench 4.0 rises to 42.3 from 31.2; ProgramBench to 29.1 from 20.3; Apex Bench v1.1 to 72.1 from 64.6.
- A third-party listing on OpenRouter shows a different Terminal-Bench 4.0 score of 26.8% for a V4.1 Flash setting, a reminder that results depend on how tests are run.
- DeepSeek says the release is faster and gives shorter answers than the model it replaces; no independent measurement of this was found.
- One estimate puts the files at roughly 510 GB, so running it locally needs data-centre class hardware.
- Through DeepSeek's own API it is served under the model name deepseek-flash.
Why It Matters
The AI race is no longer only about who has the single most capable model. It is also about who gives developers models they can download, change and run themselves. DeepSeek has built its reputation on releasing strong models with open weights, and V4.1-Flash-1017 continues that approach. The MIT licence is among the least restrictive in software: a company can build a product on the model, fine-tune it on its own data and sell the result without paying DeepSeek.
The focus of this update is agents, AI systems that carry out multi-step tasks such as running commands in a computer terminal, writing and fixing programs, or browsing for information. Terminal-Bench, ProgramBench and Apex Bench are tests of exactly those abilities. A jump of 11.1 points on Terminal-Bench between two releases a month apart would be large, but it is DeepSeek's own number. The lower third-party figure listed on OpenRouter shows why independent testing matters before businesses rely on the claim.
Shorter outputs matter more than they sound. Every word an AI model writes costs computing time and money. A model that reaches the same answer in fewer words is cheaper to run and quicker to respond, which is important for companies using AI at scale in customer support, coding tools or search.
For Indian developers and start-ups, open-weight releases like this are a practical option. Teams that cannot afford to train large models from scratch, or that want to keep sensitive data on their own servers, can start from openly licensed weights. The size of the model, however, means most will use it through cloud providers or APIs rather than on their own machines, and companies will also weigh questions about using a model built in China for sensitive work.
| Benchmark (DeepSeek's figures) | V4.1-Flash (September) | V4.1-Flash-1017 |
|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| ProgramBench | 20.3 | 29.1 |
| Apex Bench v1.1 | 64.6 | 72.1 |
Impact
Short-term: Developers can download the new weights immediately and compare them with the September release on their own tasks. Independent benchmark groups are likely to publish their own scores, which will show whether DeepSeek's gains hold up.
Long-term: Frequent open releases from DeepSeek keep pressure on closed-model companies on price and openness, and give smaller firms around the world, including in India, a capable base to build on without large licensing costs.
Who is affected: AI developers and start-ups building agents and coding tools; companies choosing between open and closed AI models; and rival AI labs competing on cost and capability.
Key Takeaway
DeepSeek's V4.1-Flash-1017, released on 17 October 2026 with MIT-licensed open weights, claims sharp gains on agent benchmarks and shorter answers, though the company's numbers still need independent confirmation.
Questions and Answers
What is DeepSeek V4.1-Flash-1017?
It is an updated snapshot of DeepSeek's V4.1-Flash model, released on 17 October 2026, aimed at better performance on agent tasks and shorter responses. Its weights are on Hugging Face under the MIT licence.
Is DeepSeek V4.1-Flash-1017 free to use commercially?
Its weights are published under the MIT licence, which allows commercial use, modification and redistribution. Using it through DeepSeek's own API is a paid service.
How much better is it than the earlier version?
DeepSeek's own tables show Terminal-Bench 4.0 rising from 31.2 to 42.3, ProgramBench from 20.3 to 29.1 and Apex Bench v1.1 from 64.6 to 72.1. A third-party listing shows a lower Terminal-Bench figure, so independent tests may differ.
Can I run DeepSeek V4.1-Flash-1017 on my own computer?
Practically, no. One estimate puts the model files at around 510 GB, which needs powerful data-centre hardware. Most users will reach it through cloud services or DeepSeek's API.
Disclaimer: Prepared by the Peepals newsroom from publicly available sources with AI assistance. Information is accurate to the best of our knowledge at the time of publication and may change. Images may be representative. Not professional advice. Report an error via our contact page.
Sourced and fact-checked by the Peepals Global Editorial Team
Reported, fact-checked and published by the Peepals Global Editorial Team.
