DeepSeek-V4.1-Flash is an open-weight model from DeepSeek that reads text and images, can take in up to a million tokens, and is built to make long agent tasks cheaper.
In plain words
It is a mixture-of-experts model: of its 552 billion main parameters, only 8 billion work on each token while it reads the input, and 16 billion while it writes. Its key-value cache, the notes a model keeps about text it has already read, takes 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash. You can set how hard it reasons on a scale from 1 to 100. DeepSeek says a later update, DeepSeek-V4.1-Flash-1017, scores better and finishes tasks faster. The weights are under the MIT licence.
A quick example
Imagine a coding agent that must read a whole large project before it changes one file. A model that uses few parameters while reading, and keeps compact notes about what it read, makes that long reading step cheaper.
Where you'll meet it
- Coding agents that read a lot of code and documents before they act.
- Questions that mix pictures and text in very long inputs.
Related
Sources
- DocsDeepSeek-V4.1-Flash model card, DeepSeek (Hugging Face) · read 17 Oct 2026
- DocsDeepSeek-V4.1-Flash-1017 model card, DeepSeek (Hugging Face) · read 17 Oct 2026
Checked by WhatIsAI against the sources above · 17 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.