On October 20, 2026 (about October 20–21 Beijing time), the FastVideo team at UCSD’s Hao AI Lab—working with Nuva Lab—released FastH3 Omni: a distilled MiniMax H3 model for reference-to-video-and-audio (Ref2VA). A prompt plus ordered image, video, and voice references can produce a finished clip with synchronized sound, typically in the 5–15 second request range the release demos. Weights and ComfyUI-oriented paths are public; the practical questions for local users are whether a given GPU can run it, what the accelerate numbers actually measure, and whether commercial use is allowed under the inherited license.
What the release actually ships for local boxes: FastVideo documents NVFP4 QAD recipes for NVIDIA DGX Spark (GB10, 128 GB unified memory) and RTX 5090 (32 GB). Both device profiles default to 480p. Spark keeps model components resident; the RTX 5090 profile offloads the text encoder and needs a precomputed AdaLN table before inference. For hardware without NVFP4, the project points to FP8 weights and device profiles, including an RTX 4090 path with layerwise offload, plus ComfyUI FP8 checkpoints. The bundled NVFP4 text encoder needs Blackwell silicon; other GPUs need compatible BF16 encoder components. NVFP4 QAD and FP8 are separate numerical variants—their outputs can differ from the BF16 Hugging Face release. Optional NVFP4 VAE decoders offer a 36-layer quality tier and an 8-layer fast tier; those local decoder choices have their own memory and precision settings, so they do not inherit the eight-GPU timing table below. Server-side serving is a different track (UniServe on multi-GPU cohorts); this article stays on the local and license questions.
Read the official speed claim with its conditions attached. Across 27 matched workloads on eight GB200 GPUs, FastH3 Omni shows 5.6–11.9× lower end-to-end latency than Base MiniMax H3 Ref2VA. The comparison covers full serving recipes (reference encoding, generation, decoding, finished MP4) at a 1344×768 canvas, with Base on 50 transformer evaluations and FastH3 Omni on eight. That figure is not a DigIntoAI benchmark, and it is not a claim that a single home GPU is “about 12× faster.” Local 480p recipes, offload settings, and quantized variants will not match those eight-GB200 numbers.
License is the other half of “can I ship this.” Use of the FastH3 Omni weights is governed by the MiniMax-H3 Community License Agreement. Applicable Territory is worldwide excluding the Excluded Territories: the European Union, the United Kingdom, the Republic of Korea, and the United States of America. Use, distribution, or display of the works or their outputs outside that Applicable Territory is not authorized by the community terms. Separately, if commercial products and services that use MiniMax H3 or MiniMax H3 Works generate more than 20 million US dollars (or equivalent) in yearly revenue, the licensee must obtain prior written authorization from MiniMax (contact api@minimax.io with the subject line specified in the license). The community grant is therefore not “free commercial use worldwide.” We have not run FastH3 Omni locally and have not audited any specific product for compliance—read the full license text before deploying.
For related MiniMax open-weight context see MiniMax M3.1 Flash open weights on Hugging Face. Another single-card local video stack on Blackwell consumer silicon is covered in EdenLive on one RTX 5090. For reference-driven shot control in a closed product stack, compare Kling 4.0 keyframes and references.