About FastH3 8-Step V2
FastH3 8-Step V2 is FastVideo's distillation of MiniMax H3 (Hailuo 03): DMD2 step distillation plus a trained Video Sparse Attention policy, so it generates the same text-to-video with native audio in 8 sampling steps instead of 28.
Fine-tuning trains a LoRA against FastVideo's de-distillation training adapter so the adapter keeps the 8-step schedule. Text-to-video only: it has no first-frame or reference conditioning. The checkpoint is the int8 ConvRot DiT from the FastVideo Comfy repack, sharing MiniMax H3's text encoder and VAEs.
| Metric | Value |
|---|---|
| Parameter Count | 33B (plus 32B encoder) |
| Mixture of Experts | No |
| Sampling Steps | 8 |
| Native Audio | Yes |
Ready to build with FastH3 8-Step V2?
Try FastH3 8-Step V2 in the Workbench to prompt it, compare outputs, and iterate on prompts without writing any code. When you're ready to ship, call the same model from our API and build your own apps on top of it.
curl -sSf -X POST https://hub.oxen.ai/api/ai/videos/generate \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OXEN_API_KEY" \
-d '{
"model": "fasth3-8-step-v2"
}'API endpoint
See the API reference for request and response formats.
https://hub.oxen.ai/api/ai/videos/generate