DeepSeek released DeepSeek-V4.1-Flash, an open-weight model that reads both images and text. The weights are published under the MIT licence. It holds up to one million tokens of context. The design cuts memory used to store context to 890 bytes per token. That is roughly a quarter of the previous DeepSeek-V4-Flash. Only 8 billion parameters run when reading input and 16 billion when writing output. Users can set reasoning effort on a scale from 1 to 100. DeepSeek reports scores above Opus-5.0 and GPT-5.6 on several coding and agent tests. It trails them on some harder tests, including Terminal Bench 3.0 and 4.0. The release ships no standard chat template, only reference code for building prompts.
What changed
DeepSeek-V4-Flash used about four times more memory per token of context.
What it unlocks
Running a million-token, image-and-text agent model locally under an MIT licence.
- 552B backbone, 8B/16B active
- 890 bytes KV cache per token
- 1M-token context window
- 45T training tokens
What you need to act on it
- weights downloaded and converted
- substantial GPU hardware
- own inference setup; no hosted provider yet
Sources