README.md

---
license: other
license_name: minimax-h3-community
license_link: LICENSE
base_model:
  - FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
tags:
  - text-to-video
  - audio
  - apple-silicon
  - metal
  - int8
  - fasth3
  - vsa
---

# FastH3 VSA · H3ddle INT8 ConvRot

This repository contains the learned-VSA FastH3 Preview v1 transformer converted
for H3ddle's native Apple-Silicon Metal engine. It is published with permission
from the MiniMax/Hailuo team and remains subject to the included MiniMax H3
Community License Agreement.

## Use in H3ddle

**FastH3 · VSA is included in
[H3ddle](https://github.com/AlexanderIstomin/h3ddle)'s managed model catalog.** Users do not
need to download this repository manually, merge checkpoints, run a conversion
script, or configure a model folder. In H3ddle, open **Models**, choose
**FastH3 · VSA** under **Video**, and select **Install**. The app downloads and
verifies this pinned transformer and reuses shared H3 components that are
already installed.

The source checkpoint is
[`FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree`](https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree)
at revision `b65818d41939b5085451074fe8ca8b799f8d4921`.

## Package

`diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors`

- Size: `22,966,486,018` bytes
- SHA-256: `53335dfb80a09e9de0ef8dce2ef36b5b91e9167044b4da7a3a1219746f474f40`
- H3ddle profile: `fasth3`
- Serving schedule: exactly four calls
- Attention: learned VSA, format 2, tile size 64, sparsity 0.9
- Conditioning: text-to-video with jointly generated audio only

The conversion keeps H3ddle's input-major INT8 ConvRot transformer core,
translates Diffusers' value-first SwiGLU layout to the native engine's
gate-first layout, preserves the trained four-call AdaLN function as exact
lookup rows, and includes all 50 learned VSA compression-gate projections.

This file is not a Diffusers checkpoint. Use it with H3ddle/h3.c and the shared
MiniMax H3 text encoder, video VAE, audio VAE, tokenizer, and configuration
files expected by the app's managed model manifest.

## Apple Silicon performance

A controlled 512×512, 124-frame comparison on a 32 GB M1 Pro used the same
prompt, soundscape, seed, and cache settings. The measured outputs were
5.175-second H.264 clips at 24 fps with stereo AAC:

| H3 model | Released passes | Attention | End-to-end time | Compared with VSA | Evidence |
| --- | ---: | --- | ---: | ---: | --- |
| Standard | 20 | Dense | ~4,253 s (~70.9 min) | ~6.11× slower | Schedule-normalized estimate |
| Turbo | 8 | Dense | ~1,782 s (~29.7 min) | ~2.56× slower | Schedule-normalized estimate |
| FastH3 Dense | 4 | Dense | 958.174 s (16.0 min) | 1.377× slower | Measured conversion baseline |
| **FastH3 · VSA** | **4** | **Learned VSA, 90% sparse** | **696.043 s (11.6 min)** | **Baseline** | **Measured** |

The Standard and Turbo rows are estimates, not completed matched runs. They
subtract the 134.439-second fixed pipeline cost observed in a phase-instrumented
matched-shape receipt, scale only the measured dense transformer work from four
passes to the released 8- and 20-pass schedules, and then add the fixed cost
back. The FastH3 Dense and VSA rows are direct end-to-end measurements.

These are local M1 Pro results, not projections for every Apple-Silicon device.
The direct four-call comparison is a 27.4% end-to-end reduction, or 1.377×
throughput, for learned VSA versus Dense.

The native deterministic VSA GPU/CPU test measured cosine similarity
`0.999998633`, mean absolute error `0.000065549`, and maximum error
`0.000706643` after BF16 output rounding.

## Limitations

FastH3 Preview v1 supports 124–362 frames and requires a short edge of at least
480 pixels. It does not support still-image generation, start/end frames,
ordered image references, or video inpainting. H3ddle enforces these constraints
in both the app and the native engine.