README.md

---
license: other
license_name: minimax-h3-community-license
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
base_model: MiniMaxAI/MiniMax-H3
base_model_relation: quantized
pipeline_tag: text-to-video
tags:
  - nvfp4
  - quantization
  - modelopt
  - comfyui
  - sparse-attention
  - video-generation
  - audio-generation
---

# MiniMax-H3 experiments

Measured experiments on making [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) (video + audio) smaller and
faster on consumer Blackwell GPUs, with files that load in stock ComfyUI. Quantization uses
[NVIDIA Model Optimizer](https://github.com/NVIDIA/Model-Optimizer); every number comes from a results file and is
regenerated by one script.

> **Work in progress.** Results are final for what they cover; new experiments are added as they finish.

## Contents

| Page | What it covers |
|---|---|
| [Calibrated NVFP4 checkpoint](docs/nvfp4.md) | 12.7 GB NVFP4 transformer, recipe sweep, comparison with published NVFP4 files, RTX 5090 speed, comparison videos |
| [Attention config LoRAs](docs/attention.md) | per-layer kitchen INT8 / SOL sparse attention allocations that beat the reference config LoRA |
| [AdaLN analysis](docs/adaln.md) | why 13B of the 33B parameters reduce to a rank-12 timestep basis ("pruned" checkpoints) |
| [TaoMate-H3 FL2VA LoRA](docs/taomate.md) | TaoLive's 3-step FL2VA LoRA with keys converted for stock ComfyUI, and the sampling schedule from their released code |
| [Methods](docs/methods.md) | how quality is measured (one-step error), noise floor, simulation vs ComfyUI kernels |

## Headline results (RTX 5090, 960×544, 5 s clip)

| Setup | Transformer | s/step (20 steps) | Video / audio error |
|---|---|---|---|
| ComfyUI repack, INT8 ConvRot | 21.0 GB | 3.87 | – |
| **This repo: calibrated NVFP4** | **12.7 GB** | **3.53** | 15.4% / 15.3% |
| … + attention config LoRA | 12.7 GB | **1.76** | |
| lilcheaty `pruned_nvfp4` (uncalibrated, same layout) | 12.5 GB | – | 18.2% / 16.8% |
| rockerBOO `pruned_nvfp4` (FP8 attention, BF16 edges) | 20.1 GB | 4.14 (ConvRot variant) | 12.7% / 9.1% |

Errors are one-step output error against BF16 on held-out prompts (see [Methods](docs/methods.md)); the attention
LoRA's own error is measured separately on the [attention page](docs/attention.md).

## Key findings

- **13 of H3's 33.1B transformer parameters are AdaLN projections that depend only on the timestep** and factor
  exactly into a rank-12 basis (66 → ~40 GB in BF16) at any step count or shift. ComfyUI's pruned build does the same
  at rank 8; at the output the two are indistinguishable.
- **4-bit weights, not 4-bit activations, carry most of the NVFP4 error**; calibration (local Hessian) cuts the error
  of a naive cast by ~15% at the same size and speed. Block-wise scale learning overfit and did not help.
- **On an RTX 5090 attention is ~73% of each step**, so NVFP4 alone is ~9% faster than INT8; with an attention
  config LoRA it is ~2× faster per step than plain INT8.
- **Our attention allocations beat the reference config LoRA at the same speed**: ~18% less video error and ~15% less
  audio error on the 4-step stack they were tuned on, and still better on the 20-step base stack.

## Files

| File | What | Size |
|---|---|---|
| `diffusion_models/minimax_h3_fl2va_pruned_nvfp4_calibrated.safetensors` | FL2VA / T2VA transformer, all block GEMMs NVFP4 (local-Hessian), AdaLN curve form rank 12 | 12.7 GB |
| `loras/minimax_h3_attn_balanced.safetensors` | attention config LoRA: reference speed, less error | ~10 KB |
| `loras/minimax_h3_attn_quality.safetensors` | attention config LoRA: ~4% slower, clearly less error | ~10 KB |
| `loras/minimax_h3_attn_fast_dmad4step.safetensors` | attention config LoRA for the dmad 4-step stack | ~10 KB |
| `loras/sinks/*.safetensors` | attention LoRAs with audio/text sinks, including ref2va 768×1344 ones that also suit the SMD 8-step LoRA (x6: ~2.8× faster sampling); need the ComfyUI PR in [Attention](docs/attention.md) | ~10 KB |
| `distill_loras/taomate_h3_fl2va_step300_bf16.safetensors` | [TaoMate-H3-FL2VA](https://huggingface.co/TaoLiveAIGC/TaoMate-H3-FL2VA) 3-step weight LoRA (step 300, rank 128), converted to ComfyUI keys; see [TaoMate](docs/taomate.md) | 1.24 GB |

**Checkpoint:** put it in `models/diffusion_models/` and pick it in the H3 template's **Load Diffusion Model** node;
keep the usual H3 text encoder and VAEs. "Pruned" means the 13B AdaLN parameters are replaced by an equivalent
timestep table plus rank-12 projections; nothing else is removed. NVFP4 runs natively on Blackwell (RTX 50-series,
B200); other GPUs load it but dequantize. Weight LoRAs (e.g. 4-step distillation) are merged by ComfyUI by
re-quantizing with fresh scales, which discards the calibration; the bypass LoRA loader keeps it at ~15% speed cost.

**TaoMate LoRA:** a weight LoRA, unlike the attention LoRAs below. Load with **Load LoRA
(Model Only)** at strength 1.0, before any attention LoRA; sampling settings are on the [TaoMate](docs/taomate.md) page.

**Attention LoRAs:** load with **Load LoRA (Model Only)** at strength 1.0, after any weight LoRA, with attention
override nodes disabled. They hold no weights.

## License and attribution

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights
Reserved. These files are derivatives of MiniMax-H3 and are distributed under the same agreement, including its
Acceptable Use Policy and territorial restrictions; see `LICENSE.MiniMax-H3`, `NOTICE` and the
[license](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE). The H3 text encoder is Qwen3-VL-32B
(Apache 2.0). Not affiliated with or endorsed by MiniMax or NVIDIA.