What actually shipped
Until November 21, 2026, pi-H3 was a demo you could watch and a set of claims you could read. The November 10 announcement showed adapters that generate MiniMax H3 video in two or four sampling steps, with a decoder-free variant that skips the video VAE decoder entirely, predicting pixels directly at the last step.
Version 1.0, dated November 21 in the project's own release list, adds the model weights, the inference code and a ComfyUI extension, so the adapters can now be run locally instead of only tried in a hosted demo. Four adapters are published: 2-step and 4-step, each with a decoder-free DWD variant and a conventional VAE-decoder variant. The authors recommend the DWD pair. INT8 ConvRot versions of those two are provided for ComfyUI.
That is a real release, and for local H3 users it is the useful part of the story. Pika's H3 Flex, covered here in October, needed six steps and sat on roughly 210 GB of base components. pi-H3 still needs the H3 base — these are adapters, not a standalone model — but it cuts the step count to two and, in the DWD variants, removes a whole stage from the pipeline.
The version the numbers belong to
Read the release notes in order and a gap appears. Version 1.0 changed the part of the system that produces the final pixels: the convolutional refiner from the DWD paper, which the authors say used a lot of VRAM and made fine-detail motion less smooth, is replaced by a transformer refiner that keeps eight of the 24 blocks of the Sol-H3 VAE decoder and is tuned end to end.
Then, one line below: "pi-H3 0.9 (Nov 10) is the initial demo, used for all videos and benchmarks in this post."
So the 3.8-second figure, the 2.4–2.7x speedup over Sol-H3 at two steps, the side-by-side quality comparisons against LightX2V, FastH3 and other few-step H3 variants — all of it describes the build with the refiner the authors have just replaced. Version 1.0 is described as better and slightly faster, but it is not the version with published measurements. Anyone quoting pi-H3's speed today is quoting a model that is no longer the one on offer.
The hardware the speed belongs to
The headline timing is end to end, which is to the authors' credit: text encoding through to decoded video and audio, median of seven requests. It is also on a single NVIDIA B200, a data center card, and it excludes model loading, compilation, warm-up and MP4 encoding. On an H100 80GB, the authors' own per-stage breakdown puts the 2-step total at 6.19 seconds for the same 5-second 768p workload the B200 chart renders in 3.81 seconds.
Two more limits are stated by the authors and worth keeping attached to the story. DWD on H3 is, in their words, still experimental: H3 compresses 12,288 pixel values into each token, and they say they have not had the compute to fully optimize for it. And the fidelity comparisons in the post were run, per the captions, with the native H3 VAE decoder in the shared inference stack — they describe the adapters decoded conventionally, not the decoder-free output, which has its own separate comparison section.
What to do with it
If you run H3 locally, start with the DWD adapters — the pair the authors recommend. Treat the published speed as a direction, not a specification, until someone publishes 1.0 timings, on any hardware.
The license travels with the base model. pi-H3 is a derivative under the MiniMax H3 Community License, the same agreement ALFW read in October: its operative restriction is on displaying the outputs, which the license limits to its applicable territory. Running the adapters locally is not what the clause governs; publishing the output everywhere is where the terms need reading.
Sources
- [1] pi-H3 project blog, Hansheng Chen et al. (release list, benchmarks, refiner notes)Read source
- [2] pi-H3 model card (adapters, INT8 ConvRot builds, license), Hugging FaceRead source
- [3] ALFW prior coverage, for distinction: Pika H3 Flex, October 29, 2026Read source