DeepSeek reissues its MoE with FP8 weights
The MIT-licensed mixture-of-experts model returned on 8 August in an updated build shipping FP8 weights for cheaper inference. Same licence, same architecture, lower serving cost — which is a distribution decision, not a research one.
DeepSeek's MIT-licensed MoE came back in an updated build carrying FP8 weights. No new capability claim accompanies it. The change is numeric precision, and the purpose is that the model costs less to serve.
This is a more consequential kind of release than it looks. An open-weight model's real constraint is not whether you may run it but whether you can afford to. Shipping FP8 alongside the licence removes a quantisation step that every serious deployer was performing anyway — and removes the variance that comes from everyone performing it differently.
The strategic read is that DeepSeek is competing on cost of ownership rather than benchmark position. That is consistent with the rest of its year, which has included price movements in both directions on its hosted tiers. A lab that raises hosted prices while lowering self-host cost is telling you where it thinks the durable business is.
For teams evaluating open weights, FP8-at-release is now a thing to check for. Its presence means the lab has thought about your serving bill. Its absence means you will be doing that work yourself.
The Open Weights — Open-source AI, tracked daily → · Thunder Compute — Best Open Source LLMs (August 2026) →