MiniMax H3 Ref2VA in FP8: video with sound on one R9700

Send a prompt and 1 to 9 reference pictures (or one reference video), and get back a 5.2-second 864ร—480 clip with stereo sound in 30 seconds, on one Radeon AI PRO R9700 (32 GB, 300 W).

These are the weights for the native MiniMax H3 Ref2VA runner in paiton-vllm-plugin (v1.2). The runner is compiled HIP code for RDNA4, with no PyTorch, diffusers or ComfyUI inside. You call it over HTTP or from a ComfyUI node.

Numbers

The original BF16 release is 144 GB, and its transformer alone (66.3 GB) is twice the memory of the card. This build needs 50.9 GB on disk for one preset and peaks at 24.7 to 26.8 GiB of VRAM.

Time per request on one R9700 with the 4-step preset (Turbo4), 124 frames at 24 fps, from the prepared inputs to the finished MP4. v1.1.0 is our earlier ComfyUI build with a W4A8 denoiser (EliovpAI/MiniMax-H3-W4A8-Paiton-RDNA4) on the same card.

Request v1.2 (these weights) v1.1.0 Faster by
864ร—480, 1 picture 30.4 s 50.6 s 1.7ร—
864ร—480, 3 pictures 33.4 s 85.3 s 2.6ร—
864ร—480, 9 pictures 44.3 s 178.7 s 4.0ร—
1344ร—768, 9 pictures 114.1 s 258.4 s 2.3ร—
864ร—480, 1 video with its soundtrack 203.8 s not measured
864ร—480, 1 picture, 8 steps (Turbo8) 51.4 s 109.2 s 2.1ร—

Peak VRAM for one picture at 864ร—480 went from 30.6 GiB (v1.1.0) to 24.8 GiB. The server loads the weights once when it starts (76 s in our test, including a SHA-256 check of every file) and keeps them in memory between requests.

Quality, compared with the BF16 model in diffusers (MI355X, same prompts, references, seed and starting noise). CLIP-T scores how well the clip matches its prompt:

Request CLIP-T, BF16 CLIP-T, these weights
864ร—480, 1 picture 29.52 29.77
864ร—480, 3 pictures 34.68 34.35
864ร—480, 9 pictures 26.31 25.70
1344ร—768, 9 pictures 25.80 25.71
864ร—480, 1 video 29.87 29.86

Over 15 clips (two seeds, both presets) the frames have a mean similarity of 0.932 (CLIP-I) and 0.871 (DINOv2) to the same frames of the BF16 clip, where 1.0 is the same image. With nine references the subjects are all there but can stand in other places than in the BF16 clip. Method and limits: BENCHMARKS.md.

Run it

git clone --depth 1 https://github.com/Eliovp-BV/paiton-vllm-plugin.git
cd paiton-vllm-plugin
PAITON_DETACH=1 ./models/MiniMax-H3/native/launch.sh http
docker logs -f paiton-minimax-h3-v12        # wait for "READY http://..."; the first start downloads the weights
curl -o panda.mp4 -F image=@panda.png \
  -F prompt='The red panda from <Picture 1> walks through a sunlit forest. Audio: birds and a stream.' \
  http://127.0.0.1:30000/v1/videos/generations

The runner downloads these files on first use and checks each one against the SHA-256 in its checkpoints.lock.json, so you don't need to download anything by hand. For ComfyUI, copy the paiton_h3v_client node from the plugin into custom_nodes; it sends the job to the server, so ComfyUI itself can run on the CPU.

You need one R9700 with nothing else on it, Docker, 51 GB of disk for one preset, and about 41 GB of free host RAM for the fastest text-encoder path (with less, the runner reads the text encoder from disk instead).

Files

The files only work with this runner: they are packed for its kernels, not for diffusers or ComfyUI.

Folder Content Size
weights/turbo4/ transformer, 4 steps (Turbo LoRA merged, FP8) 20.3 GB
weights/turbo8/ transformer, 8 steps (Turbo LoRA merged, FP8) 20.5 GB
dec8.bin, enc.bin, audio.bin video VAE decoder (INT8), video VAE encoder (BF16), audio VAE (FP32) 3.4 GB
model/tokenizer/ tokenizer 7 MB

The text encoder (Qwen3-VL-32B, layers 0 to 49 in FP8, and the vision tower, 27.2 GB) is the same as the one of MiniMax H3 Image, so it is not stored twice: the runner downloads it from the te/ folder of EliovpAI/MiniMax-H3-Image-FP8-Paiton-RDNA4, pinned by revision and SHA-256. The changes to each file are in MODIFICATIONS.md.

License

MiniMax H3 Community License (LICENSE, NOTICE). The grant excludes the European Union, the United Kingdom, the Republic of Korea and the United States, and commercial use above the agreement's revenue threshold needs MiniMax's written authorization. Read the agreement before you download or use these files.

Downloads last month
64
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for EliovpAI/MiniMax-H3-FP8-Paiton-RDNA4

Finetuned
(239)
this model