We run diffusion and video models fast, and we do it on other people's weights. Most of what is in this org is a by-product of that work, published because it is more useful outside our cluster than inside it.
FP8 (e4m3), INT8, NF4 and precision-cast variants of open models — FLUX.1-dev, Qwen-Image, Qwen-Image-Edit, Wan 2.1, MAGI-1, LongCat-Video-Avatar, Ovi, MiniCPM-V. These are the exact artifacts our fleet loads.
Every card states which tensors were actually changed and which were left in
their original dtype, because "quantized" on its own tells you nothing about
whether the text encoder came along for the ride. Each one is linked to its
upstream repository through base_model, and inherits the upstream license —
the FLUX derivatives are non-commercial, the Qwen and Wan derivatives are
Apache-2.0.
→ Quantized weights we publish
Bring-your-own-key demos for the models we host, and static reference pages for
the open-weight ones. The interactive Spaces ship no API key: you paste your
own, it is sent only to api.wavespeed.ai, and generations are billed to your
account rather than ours.
→ Text to image · Image editing and enhancement · Video generation
Nothing here is a model we trained. The collection below points at the repositories whose weights actually run behind our endpoints — start there if you would rather run them yourself.
→ Open weights behind our endpoints
Hosted inference and docs: wavespeed.ai