# DeepSeek V4.1-Flash-1017 lifts benchmarks under an MIT licence

> DeepSeek released MIT-licensed V4.1-Flash-1017, improving on its Flash model across reasoning, coding and agent benchmarks.

*DeepSeek gives away a cheaper-to-run Flash update with a 1M-token context, though Opus 5 still leads on most listed tests.*

By Bez · InsidersFeed
Canonical: https://insidersfeed.com/news/deepseek-v4-1-flash-1017-lifts-benchmarks-under-an-mit-licence

DeepSeek is pushing down the cost of capable open-weight models again, and it is doing so without a licence fee. The company has released V4.1-Flash-1017 on Hugging Face, with the repository and weights under the MIT licence.

The [model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) calls it a minor update to V4.1-Flash, with "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks", in DeepSeek's words. It also says the model cuts task execution time and output length, which lowers cost. The card gives no figures for those savings.

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 benchmark gains over the original Flash are uneven but broad. On HLE without tools, Flash-1017 scores 42.6 against 36.8. Terminal-Bench 4.0 jumps from 31.2 to 42.3, and ProgramBench from 20.3 to 29.1. Apex Bench v1.1 moves from 64.6 to 72.1. DeepSWE v1.1 rises from 74.2 to 77.0, and BabyVision from 89.6 to 92.5.

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

Deployment takes some work. There is no Jinja chat template. DeepSeek instead supplies a Python reference implementation, encoding.py, and a Rust toolkit, deepseek-recipe, for production use.

Two details are unconfirmed. The model card gives no release date; the 17 October date comes from a news roundup at aiweekly.co. That roundup also lists 552B total parameters and new Search and Fetch APIs at $0.75 per 1,000 successful requests. The model page shows 763B, and it lists no API pricing.

## The Insider view

The commercial pitch is cost per finished task, not headline intelligence. Shorter outputs and faster execution matter most to operators running agents at volume, and the biggest jumps are in agentic and coding tests: Terminal-Bench 4.0 and ProgramBench. An MIT licence removes licensing friction for anyone who wants to host the model or build it into a product.

The exposure falls on paid closed-model vendors in the mid tier, and on open-weight rivals such as K3 and GLM-5.3, which sit in the same comparison table. Opus 5 keeps its lead on most benchmarks, so the premium tier is not displaced. Watch for independent cost measurements, since DeepSeek has not quantified its savings, and for whether the conflicting parameter count and API pricing claims are resolved. The missing chat template also means the first adopters will be teams with engineering capacity.

## Key takeaways

- DeepSeek ships a free, MIT-licensed Flash update with a 1M-token context window.
- Shorter outputs and faster task execution lower running cost, according to the model card.
- Opus 5 still beats Flash-1017 on most benchmarks DeepSeek lists.

## Sources

- [DeepSeek-V4.1-Flash-1017 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) — Hugging Face (DeepSeek), 2026-10-17