The date was kept. Read the measurement before the multiple

When FastH3-Omni was [previewed on 13 October](https://alfw.ai/article/minimax-fasth3-omni-reactor-preview/), alfw.ai noted the unusual part: a preview with a date on it. The full model and training code were due on 20 October, and on 20 October they appeared. The Hugging Face collection carries a BF16 checkpoint the card calls the final release, ComfyUI variants, quantized builds and a VAE decoder, and the training code is in the FastVideo repository with a guide. Promised open releases slip often enough that landing on the stated day is worth recording plainly.

The speed story needs slower reading. The 5.6 to 11.9 times figure is real arithmetic from a published table: 27 workloads, three seeds each, timings that include reference encoding and MP4 delivery, with a downloadable CSV and JSON. That is more documentation than most launch claims provide. It is also, by the methodology section's own account, a comparison of two complete serving recipes that differ in weights, attention patterns and runtime, measured by the team that built the faster one, on eight GB200 data-centre GPUs, with the base model running a numerical policy the team itself describes as less strict than exact. None of that makes the numbers wrong. The result is a team-reported comparison of complete model-and-serving configurations, not an isolated measurement of the weights' contribution: it shows the latency achieved by the tested recipe on eight GB200 GPUs, and it does not establish that the downloaded checkpoint will deliver the same speedup under a different runtime, hardware configuration or numerical policy. The team is unusually honest about this: its gallery states that the matched examples are "not evidence of output parity", and its local guide states flatly that the GB200 timings do not describe the RTX 5090 or Spark configurations.

What you actually get locally is more modest and more useful. The NVFP4 QAD build runs reference-guided H3 generation on a single RTX 5090 or a DGX Spark, at 480p by default, with the text encoder offloaded on the 5090 and a precomputed AdaLN table. Reducing the generation path from fifty transformer evaluations to eight is a meaningful architectural change, but its practical latency benefit depends on the hardware, runtime and configuration. An FP8 route with layerwise offload keeps an RTX 4090 in the picture. That is the practical release: not a data-centre benchmark on hardware almost nobody has, but a reference-to-video model with synced audio that fits on a desk.

"Open source" is doing work in the headline

Nuva Lab's release post opens with "Open Weight. Open Source." The model card is narrower: use of these weights is governed by the MiniMax H3 Community License Agreement. That is the same licence that covers the base H3 weights. The training code is publicly available in the FastVideo repository, but that does not mean the code and the weights share a licence; each component carries its own terms. Weights you can download under a community licence are a real contribution, and the training code being public is more than most distillations offer. But this is not an unrestricted licence, and anyone building a product on these weights should read the agreement they are actually accepting, not the launch headline. For the underlying restrictions, see our earlier [MiniMax H3 licence explainer](https://alfw.ai/article/minimax-h3-license-fine-print/), which reads through the Community License Agreement's terms, including its territorial display conditions.

The second quiet fact is the tiering. Nuva's own comparison table says the production version uses more training data and extra training stages, supports native long video without stitching, and improves quality, speed and reliability over the open checkpoint. Those are the maker's claims, not an independently verified comparison. The open release and the production offering occupy different tiers: the open release makes the weights and training code available, while Nuva's commercial product is positioned as the higher-performance option. That is a legitimate business model, and the company states it openly. But buyers should distinguish the available checkpoint from the capabilities claimed for the paid version. Judge the open release as that, and it is a strong release. Judge it by the words "unlimited" and "open source" in the launch copy, and you will be disappointed by documents its own publishers wrote.

Sources

  1. [1] Hao AI Lab (FastVideo), 'FastVideo FastH3 Omni: Open Source Real-Time Minimax FastH3 Ref2VA on NVIDIA Blackwell, RTX Spark, and RTX GPUs', 20 October 2026Read source
  2. [2] FastVideo, 'FastVideo/FastVideo-FastH3-Omni-8-Step-V1' model card, Hugging FaceRead source
  3. [3] FastVideo, 'FastH3 Omni' collection, Hugging FaceRead source
  4. [4] Nuva Lab, 'FastH3-Omni release: open weights for fast omni-reference video', 20 October 2026Read source
  5. [5] Prior alfw.ai coverage: 'FastH3-Omni Is Not a New MiniMax Model. That Is the Interesting Part.', 14 October 2026Read source