Product Launches and Community Projects

Launch: DeepSeek-V4.1-Flash (DeepSeek)

Posted by Applied AI Commons team (community team) on

MIT-licensed 552B MoE with native vision, 1M context and cheaper off-peak API rates

DeepSeek's new Flash model uses a causal encoder-decoder MoE design (552B backbone, 8B/16B active) with native image input and a much smaller KV cache. Weights are on Hugging Face under MIT; it also replaces V4-Flash in the API. The release ships no Jinja chat template, and open-source inference support was still being worked on at launch.

Released 2026-09-10 by DeepSeek. Listed by the community team; we are not affiliated.

Has anyone self-hosted V4.1-Flash yet? What hardware and serving stack are you using, and how does it compare to the API?

Link: https://deepseek.com/en/news/deepseek-v4-1-flash/

3 replies

  1. Applied AI Commons team (starter reply from the community team),

    Weights are on Hugging Face under the MIT licence, which is about as permissive as model licences get.

  2. Applied AI Commons team (starter reply from the community team),

    Be ready for some setup work if you self-host: the release shipped without a Jinja chat template and open-source inference support was still in progress at launch.

  3. Applied AI Commons team (starter reply from the community team),

    The smaller KV cache is the practical win for long conversations and many concurrent users, since memory, not compute, is often what limits serving.

More in Product Launches and Community Projects