01
Few-step post-training
FastH3 reduces the original H3 denoising schedule to four transformer calls through a trained distilled path.
FastVideo × Nuva Lab
FastH3: MiniMax H3, post-trained and optimized
FastH3 is the post-trained and optimized MiniMax H3 path produced by FastVideo. It combines few-step distillation, sparse attention, and inference work for accelerated synchronized video-and-audio generation. Nuva Lab contributes production grounding and workload evaluation.
What FastH3 is
One project, not separate FastVideo and H3 offerings
MiniMax H3 supplies the base open-weight video-and-audio model. FastVideo supplies the Apache-2.0 post-training and inference framework. FastH3 is the resulting post-trained and optimized H3 path.
Nuva Lab's public work here is FastH3 production grounding: bringing real creative workloads, evaluation criteria, and end-to-end production feedback into the collaboration. FastVideo remains the owner and maintainer of its open-source framework and FastH3 artifacts.
Technical path
How MiniMax H3 becomes FastH3
02
Video Sparse Attention
The recommended Preview v1 profile applies sparse execution to eligible video-to-video attention tiles.
03
Inference optimization
FastVideo connects the trained graph to runtime, kernel, memory, and parallel execution work.
04
Production evaluation
Nuva Lab contributes workload-grounded evaluation and feedback on complete generated outputs.
Public evidence
What the FastH3 release proves
- FastH3 Preview v1 is an open-weight four-step T2VA release built on MiniMax H3.
- The published 15-second, 768p result is faster than playback on eight NVIDIA B200 GPUs.
- The release supports variable duration, resolution, and aspect ratio in one checkpoint.
- FastVideo subsequently published local paths for NVIDIA DGX Spark and Apple Silicon.
Evidence boundaryThese are public FastH3 release results, not a universal performance claim. Output quality and latency still depend on the exact checkpoint, workload, hardware, and complete media path.
Licensing
Open framework and open weights are not the same license
FastVideo is Apache-2.0 open source. MiniMax H3 and FastH3 model weights and derivatives are governed by the MiniMax H3 Community License. That license currently excludes the United States, European Union, United Kingdom, and Republic of Korea from its default applicable territory and directs interested deployers in excluded territories to seek authorization from MiniMax.
Any use of H3 or FastH3 remains subject to the applicable model terms, deployment territory, and required authorization. The FastVideo framework license does not replace the model license.
FastH3 FAQ
FastH3 questions
What is FastH3?
FastH3 is FastVideo's few-step, sparse-distilled derivative of MiniMax H3 for synchronized video-and-audio generation. The recommended Preview v1 profile uses four transformer calls and Video Sparse Attention.
Can FastH3 run locally?
Yes. FastVideo has published FastH3 paths for NVIDIA DGX Spark and Apple Silicon, in addition to data-center NVIDIA GPUs. Local does not necessarily mean real-time; latency and memory strategy vary materially by hardware.
Is FastH3 open source?
The FastVideo framework is Apache-2.0 open source. FastH3 uses MiniMax H3 weights and derivatives under the MiniMax H3 Community License, so “open weight” is the accurate model description.
How are FastVideo and FastH3 related?
FastVideo is the Apache-2.0 post-training and inference framework. FastH3 is the MiniMax H3 derivative produced through that work, including few-step distillation and sparse optimization.
Sources
Technical releases and integrations
Working with FastH3?
Contact Nuva Lab about production-grounded evaluation and collaboration.
Contact Nuva Lab