by Hao AI Lab
Distilled MiniMax H3 reference-to-video-and-audio model from Hao AI Lab, generating in 8 steps instead of 50.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
FastH3 Omni is a distilled version of the MiniMax H3 model for reference-guided video and audio generation, released by Hao AI Lab under the FastVideo project. Given a text prompt plus image, video and voice references, it generates a finished video with synchronized sound. It uses eight transformer evaluations instead of the base model's 50.
According to the vendor, end-to-end latency is 5.6–11.9× lower across 27 matched workloads on eight GB200 GPUs, compared against the base model's full serving recipe. In one example, a 5 s single-image request dropped from 35.44 s to 5.67 s, and a 15 s request from 139.69 s to 14.14 s. FastVideo UniServe is said to reach real-time generation for 768p clips of 5, 10 and 15 s on particular tasks. The vendor states the matched examples are qualitative comparisons, not evidence of output parity with the base model.
For local use, NVFP4 QAD checkpoints and device recipes target DGX Spark and RTX 5090 through FastVideo. ComfyUI support is available, and creator coverage mentions FP8, FP4 and Apple silicon variants. Weights and a UniServe deployment are linked from the announcement blog. The license was not stated in the provided text.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Acceleration method for an open reference-to-video model, cutting 50 steps to 8 with up to 12x lower latency on GB200s. Released variants include ComfyUI, FP8, FP4 and Apple silicon.
Closest neighbours by meaning, not just by category.
Model · Hao AI Lab
Open-weight 4-step sparse-distilled version of MiniMax H3 for fast text-to-video-and-audio generation on Blackwell GPUs.
Model · MiniMax
Open-weight video generation model from MiniMax, part of its H3 family alongside H3 Image.
Model · MiniMax
Open-weights MiniMax image generation and editing model producing up to ~16 MP images in 6 sampling steps.
Model · Pruna AI
Open-weight image generation and editing model from Pruna AI, distilled from MiniMax's H3 video model to run in 8 steps.
Model · Unsloth
GGUF quantizations of MiniMax-H3 video-with-audio generation model, for stable-diffusion.cpp and Unsloth.
Model · Kijai
Community-quantized W4A8 and int8 VAE checkpoints of MiniMax H3 video model for ComfyUI, by Kijai.