README.md

---
license: other
license_name: minimax-h3-community-license-agreement
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
tags:
  - checkpoints
  - minimax-h3
---

# Checkpoints

## MiniMax-H3

Mirror of [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3), in its native per-partition layout.
Each variant folder is self-contained (transformer + text encoder + video/audio VAE + tokenizer + processor), so download only the one you need.

| Folder | Upstream partition | Tasks | Size |
|--------|--------------------|-------|------|
| `MiniMax-H3/TI2V/` | `FL2VA` | T2VA (text-to-video+audio), FL2VA / I2VA (first/last-frame image-to-video+audio) | ~144 GB |
| `MiniMax-H3/Ref2V/` | `Ref2VA` | Ref2VA (reference-to-video+audio) | ~144 GB |

Layout of each variant:

```
transformer/    MiniMaxH3DiTModel, 13 safetensors shards
text_encoder/   Qwen3-VL encoder, 14 safetensors shards
video_vae/      video VAE code + source/model.safetensors
audio_vae/      audio VAE code + model.safetensors
tokenizer/  processor/  model_index.json  LICENSE
```

### Download one variant

```bash
# T2V / I2V
hf download mingyang-wu/Checkpoints --repo-type dataset \
  --include "MiniMax-H3/TI2V/*" --local-dir ./ckpts

# Ref2V
hf download mingyang-wu/Checkpoints --repo-type dataset \
  --include "MiniMax-H3/Ref2V/*" --local-dir ./ckpts
```

Only the transformer weights (skip the shared text encoder / VAE you already have):

```bash
hf download mingyang-wu/Checkpoints --repo-type dataset \
  --include "MiniMax-H3/Ref2V/transformer/*" --local-dir ./ckpts
```

Python:

```python
from huggingface_hub import snapshot_download
path = snapshot_download("mingyang-wu/Checkpoints", repo_type="dataset",
                         allow_patterns="MiniMax-H3/Ref2V/*", local_dir="./ckpts")
# -> ./ckpts/MiniMax-H3/Ref2V
```

The weights are distributed under the MiniMax-H3 Community License (see `LICENSE` inside each variant folder).