Hugging Face's logo

StillDeadcode
/
minimax-h3-i8

Text-to-Video
MiniMax H3
radiance
rocm
amd
rdna4
int8
w8a8
iq8r
hadamard
text-to-audio-video
latent-upscaler
Model card Files
xet
 Community
minimax-h3-i8 / README.md
StillDeadcode's picture
StillDeadcode
IQ8R container
f565097 verified
preview code
|
Raw History Blame Contribute Delete
9.87 kB
metadata 
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 — video with a synchronized stereo soundtrack from text — as a single .rad container for the radiance inference engine (AMD RDNA4, ROCm). It replaces 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.

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 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 

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.

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:

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 

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.