File size: 9,867 Bytes
f565097 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 | ---
license: other
license_name: minimax-h3-community-license-agreement
license_link: LICENSE
pipeline_tag: text-to-video
library_name: radiance
tags:
- radiance
- rocm
- amd
- rdna4
- int8
- w8a8
- iq8r
- hadamard
- text-to-video
- text-to-audio-video
- minimax-h3
- latent-upscaler
base_model:
- MiniMaxAI/MiniMax-H3
- LBH-123-AI/Minimax_h3_latent_Upscaler
base_model_relation: quantized
---
# MiniMax-H3, int8 — radiance container
[MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) — video with a synchronized stereo
soundtrack from text — as a single `.rad` container for the
[radiance](https://codeberg.org/StillDeadcode/radiance) inference engine (AMD RDNA4, ROCm). It replaces
[StillDeadcode/minimax-h3-fp8](https://huggingface.co/StillDeadcode/minimax-h3-fp8), with the same eight bits a
weight and no E4M3 weight left.
The MiniMax H3 Community License does not cover the European Union, the United Kingdom, the
Republic of Korea or the United States of America; see [License](#license).
| | |
|---|---|
| File | `minimax-h3-i8.rad` — 62.79 GiB |
| Weights | int8 throughout: the text encoder's linears (the 50 layers of Qwen3-VL-32B the model reads) and the token refiner's IQ8R (codes −127..127, each 128 of K rotated by the unnormalised Hadamard, a u8 scale per row and 256 of K under a bf16 scale a row); the modulation projections the same rotation with a bf16 scale per row and 128; the DiT's 50 blocks' linears unrotated int8 with a bf16 scale per row and 128; the video decoder bf16; the audio decoder f32 |
| Activations | int8 (W8A8), rotated where the weight is |
| Latent upscaler | [LBH-123-AI/Minimax_h3_latent_Upscaler](https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler) v1 (3D conv, 345M parameters, bf16; Apache-2.0) |
| Output | MP4: H.264 at 24 fps, up to 1344x768, 5.17 to 14.4 s; AAC at 32 kHz stereo |
| Needs | radiance built from main after 1.17.0 (the text encoder's, the refiner's and the modulation projections' int8 forms are not in 1.17.0 and come with the next release); two 32 GB cards and about 39 GiB of pinned host memory |
## Serve
```sh
radiance --model minimax-h3-i8.rad --tp 2 --vision-tower-placement ram --host 0.0.0.0 --port 8000
curl -s http://localhost:8000/v1/videos -H 'Content-Type: application/json' -d '{
"prompt": "A red fox trotting through a snowy pine forest, snow crunching underfoot",
"seconds": "5", "size": "1344x768", "seed": 7}' # -> {"id": "video_...", "status": "queued", ...}
curl -s http://localhost:8000/v1/videos/video_... # poll until "status": "completed"
curl -s http://localhost:8000/v1/videos/video_.../content -o fox.mp4
```
The server speaks OpenAI's video jobs API (`POST /v1/videos`, `GET /v1/videos/{id}`,
`GET /v1/videos/{id}/content`, list and delete), and SGLang's `target` field:
- `prompt` is read as written, up to 2048 tokens; the model follows long structured descriptions
(the shot, its soundscape, its music) best.
- `seconds` is rounded up to the model's length steps (17 frames): 5.17 to 14.4 s. `size` is
`WIDTHxHEIGHT`, multiples of 32, at most 1344x768 pixels of area.
- `seed` makes a video repeatable on the same deployment; `num_inference_steps` counts sigma grid
points as diffusers does (50 by default, at most 65). The model is guidance-distilled: no guidance
scale, no negative prompt.
- Text to video and audio only; the image- and reference-conditioned forms are not in this file.
On two Radeon AI PRO R9700s (gfx1201) the FP8 container's 5.17 s clip at 1344x768 takes about 8.8
minutes at the default 50 steps; a step of this container takes about 7% longer (measured below,
at 768x448). The text encoder and the modulation projections are read from pinned host memory,
which `--vision-tower-placement ram` selects. For the eight-step distillation, see
[StillDeadcode/fasth3-8-step-v2-i8](https://huggingface.co/StillDeadcode/fasth3-8-step-v2-i8).
Against the FP8 container on the same two cards, same prompt and seed, a 5.17 s clip at 768x448: 40 s against 38 s at 12 steps and 116 s against 108 s at 36, so a step of the int8 DiT takes about 7% longer; two runs of this container give the same video byte for byte, and its frames show the FP8 container's shot (PSNR 22.7 dB between the two at 12 steps). Not measured against diffusers' bf16 pipeline.
## Latent upscaling
The container carries LBH-123-AI's latent upscaler. A request with `"latent_upscale": 2` runs its
first steps at half the size (672x384 for 1344x768), the upscaler enlarges the clean latent, and the
last `upscale_refine_steps` (12 by default) run at the full size from a re-noised copy of it:
```sh
curl -s http://localhost:8000/v1/videos -H 'Content-Type: application/json' -d '{
"prompt": "...", "seconds": "5", "size": "1344x768", "seed": 7, "latent_upscale": 2}'
```
| 5.17 s at 1344x768, 50 steps, 2x R9700, the FP8 container | time |
|---|---|
| native | 8.8 min |
| `latent_upscale` 2, 12 refine steps (the default) | 3.4 min |
| `latent_upscale` 2, 20 refine steps | 4.5 min |
The factor goes up to 4, but at 1344x768 a factor of 4 starts from 352x192, which is too small for
the model to lay out a clean scene: expect soft, ghosted frames. The video object of an upscaled
video reports `latent_upscale`, `upscale_refine_steps` and `upscale_from` (the small size).
## How this file was made
```sh
rad-convert MiniMaxAI/MiniMax-H3 --recipe minimax-h3-i8.recipe -o minimax-h3-i8.rad
```
`MiniMaxAI/MiniMax-H3` is the downloaded pipeline directory; only its text-to-video-and-audio
components are read (`text_encoder/`, `transformer/`, `vae/`, `audio_vae/`, the processor, the
tokenizer and both schedulers), and the upscaler as a companion folder `latent_upscaler/` (its
`minimax_h3_latent_upscaler_3d_conv_v1/config.json` and bf16 weights, renamed
`diffusion_pytorch_model.safetensors`). It converts with the rest. The recipe (`minimax-h3-i8.recipe` in this repository) names what is
quantised; everything it does not name is the checkpoint's own dtype:
```
te.blk.*.attn_qkv.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
te.blk.*.attn_output.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
te.blk.*.ffn_gate_up.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
te.blk.*.ffn_down.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
te.blk.*.attn_norm.weight cast dtype=f32
te.blk.*.ffn_norm.weight cast dtype=f32
ref.blk.*.attn_qkv.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
ref.blk.*.attn_output.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
ref.blk.*.ffn_gate_up.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
ref.blk.*.ffn_down.weight rtn codes=i8 group=256 scale=u8 scale2=bf16 block2=1x* transform=fwht128 clamp=sym rule=search
dit.blk.*.attn_qkv.weight rtn codes=i8 group=128 scale=bf16 clamp=sym
dit.blk.*.attn_output.weight rtn codes=i8 group=128 scale=bf16 clamp=sym
dit.blk.*.ffn_gate_up.weight rtn codes=i8 group=128 scale=bf16 clamp=sym
dit.blk.*.ffn_down.weight rtn codes=i8 group=128 scale=bf16 clamp=sym
dit.blk.*.ada*.weight rtn codes=i8 group=128 scale=bf16 transform=fwht128 clamp=sym rule=search
vae.blk.*.attn_qkv.weight cast dtype=bf16
vae.blk.*.attn_out.weight cast dtype=bf16
vae.blk.*.ffn_up.weight cast dtype=bf16
vae.blk.*.ffn_down.weight cast dtype=bf16
vae.proj_out.* cast dtype=bf16
aud.*.weight cast dtype=f32
aenc.d*.w cast dtype=bf16
aenc.d*.b cast dtype=bf16
aenc.pb.* cast dtype=bf16
aenc.b*.u*.c1.weight cast dtype=f32
aenc.b*.u*.c2.weight cast dtype=f32
aenc.cin.weight cast dtype=f32
aenc.fin.weight cast dtype=f32
venc.l*.b*.c1w cast dtype=bf16
venc.l*.b*.c2w cast dtype=bf16
venc.l*.dnw cast dtype=bf16
venc.cout.weight cast dtype=bf16
ups.embed.* cast dtype=f32
ups.*.emb_layers.* cast dtype=f32
```
## License
MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax
(`LICENSE` in this repository; `NOTICE` carries its notice). The agreement grants its rights only
within its Applicable Territory: worldwide except the European Union, the United Kingdom, the
Republic of Korea and the United States of America, where it does not license using, reproducing or
distributing the model or this file. Its use restrictions and Acceptable Use Policy (Exhibit A of
`LICENSE`) apply to everyone who downloads or uses this file, and to anyone it is passed on to.
This container is a modified copy of MiniMax H3: the linears named above are re-encoded as int8, the
video decoder's linears cast from f32 to bf16, the audio decoder's weight norms folded, the text
encoder's last 14 layers (which the model does not read) left out, and everything repacked into one
`.rad` file. The text encoder is Qwen3-VL-32B, licensed under the Apache License 2.0
(`LICENSE-Qwen3-VL`). The latent upscaler is LBH-123-AI's Minimax_h3_latent_Upscaler v1, licensed
under the Apache License 2.0 (`LICENSE-latent-upscaler`); its weights are included as published,
its embedding MLP cast from bf16 to f32.
|