EliovpAI/MiniMax-H3-W4A8-Paiton-RDNA4

MiniMax H3 W4A8 with Paiton on Radeon AI PRO R9700

Runtime package and compiled artifacts; no model weights are mirrored here. This package executes upstream quantized MiniMax H3 FL2VA with H3 Turbo on one 32 GB gfx1201 GPU. Original checkpoints download from their pinned publishers. The Paiton runtime bindings use Apache-2.0; H3 and its derivatives retain the custom community license above. See the complete third-party notices.

Create video with native stereo audio locally on one Radeon AI PRO R9700, 32 GB. The included workflow generates a continuous 15.08-second, 864×480 clip at 24 fps. Powered by MiniMax H3.

Paiton takes 332.85 seconds versus 399.40 seconds for the matched stock request: 16.66% lower end-to-end latency and 19.99% more clips per hour. Fresh conditioning, video/audio decoding and writing a playable MP4 are included. All three paired long clips have equal video/audio latents and byte-identical MP4s.

Start ComfyUI with one command

git clone --depth 1 https://github.com/Eliovp-BV/paiton-vllm-plugin.git && cd paiton-vllm-plugin && ./models/MiniMax-H3/launch.sh

Open ComfyUI. The first visit opens the connected Turbo8 15-second workflow. Edit the prompt and click Run. Choose paiton or stock in MiniMax H3 Engine. Keep the seed fixed when comparing engines.

The helper pulls the versioned GHCR artifact image, prepares pinned ComfyUI components locally, verifies the model downloads and starts the interface. Models and caches persist in ~/.local/share/paiton/minimax-h3/ (or your XDG data directory); videos go to ~/paiton-videos/. Both paths are independent of the working directory. Subsequent launches reuse the local image and model cache. Ports bind to localhost.

From this model directory, ./launch.sh --logs follows startup and ./launch.sh --stop stops H3 while preserving caches, videos and user settings. PAITON_H3_DATA, PAITON_H3_OUTPUTS and PAITON_H3_PORT override those defaults. Stop other GPU generation workloads before starting H3.

With weights already present, the first 15-second requests took 423.6 s stock and 381.0 s Paiton, after process setup. First download, verification, local image preparation and a cold disk/runtime cache add time. One 32 GB R9700 is required. Long-clip sampled driver memory reached 30.34 GiB. We tested 16 GB system RAM with 4 GB swap; lifetime process RSS reached 12.09 GiB, and up to 1.24 GiB of process swap was sampled. 24 GB or more host RAM is recommended for the browser, longer requests and other applications. Allow 60 GB free disk for the approximately 6 GB image, 33.96 GB selected weights, caches and outputs. Both adapters share the large components and total 35.92 GB of checkpoint files. Use SSD storage. Filesystems without hard-link support can require another 34 GB for duplicate cache copies.

Choose a workflow

Workflow	Output	Purpose
turbo8-15s.json	362 frames, 15.0833 s	Default continuous scene; eight evaluations
turbo8.json	124 frames, 5.1667 s	Shorter eight-step scene
turbo4.json	124 frames, 5.1667 s	Faster four-step alternative with a quality tradeoff

All use 864×480, 24 fps and native 32 kHz stereo. Find them in the included workflow templates or load the JSON from comfyui/workflows/. H3 snaps frame counts upward to its 17k+5 grid: requesting 360 frames produces 362. It does not produce exactly 15.000 seconds at this setting.

To add the faster short preset, run ./run.sh download --profile turbo4 and open turbo4.json. To install only Turbo4 initially, use PAITON_H3_PRESET=turbo4 ./launch.sh and open the URL printed by the launcher.

Terminal generation

Stop the interface before launching a separate GPU pipeline:

./launch.sh --stop
./run.sh download --profile turbo8
./run.sh generate --engine paiton --preset turbo8 --frames 362 \
  --prompt 'A small red fox trots through a sunlit woodland clearing, pauses beside a shallow stream, and looks toward the camera. Natural realistic fur, continuous gentle camera movement, coherent anatomy. Audio: birds chirping, leaves rustling and quiet flowing water; no music and no speech.' \
  --seed 771 --output /outputs/fox-15s

The clip is saved as ~/paiton-videos/fox-15s/clip-0.mp4. For a faster short request use --preset turbo4 --frames 124 after downloading its adapter. generate includes startup for each new process; benchmark mode keeps the pipeline loaded.

Reproduce the comparison

./launch.sh --stop
./run.sh benchmark --engine stock --preset turbo8 --frames 362 --output /outputs/stock-15s
./run.sh benchmark --engine paiton --preset turbo8 --frames 362 --output /outputs/paiton-15s
./run.sh report /outputs/paiton-15s
./run.sh inspect /outputs/paiton-15s/clip-2.mp4 --output /outputs/paiton-15s/quality

Each benchmark performs one warmup and two measured requests, encoding the prompt afresh every time. It retains commands, MP4s, raw timings, latent tensors and telemetry. Use fresh output directories and run the engines sequentially with the same settings. ComfyUI's node cache may reuse identical requests, so cached UI timings are a separate scenario.

Matched Turbo8, 15.08-second clip	Stock	Paiton
Complete playable MP4	399.40 s	332.85 s
Pipeline before file encoding	385.60 s	318.98 s
Fixed-length clips per hour	9.01	10.82
Generated video seconds per wall second	0.0378	0.0453

The earlier three-prompt qualification also measured the short presets:

Matched preset, 5.17-second clip	Stock	Paiton	Lower latency
Turbo4, four evaluations	64.15 s	54.38 s	15.24%
Turbo8, eight evaluations	95.83 s	80.60 s	15.89%

Full settings, component times, memory definitions, per-run data and quality limits are in BENCHMARKS.md.

Requirements and supported scope

• Linux x86-64, Docker Engine with Compose, and working /dev/kfd and /dev/dri access for gfx1201.
• Pinned Torch 2.12.0+rocm7.14.0, Triton 3.7.1 ROCm, ComfyUI commit efa6c8f804bff78b46a0fd458ebd2e47bba07a30, kitchen 0.2.33 and AIMDO 0.5.2. The container setup supplies the software; the host supplies the AMD driver.
• A single GPU, batch one, Euler/simple, shifts video/audio 6/3, BasicGuider and exactly four or eight denoiser evaluations. The qualified long preset uses eight.
• Component scheduling uses smart memory, Dynamic VRAM, two transfer streams, --fast-disk and a 2 GB GPU reserve. Avoid --high-ram on the tested 16 GB host.

The denoiser uses upstream pruned FL2VA W4A8 weights. The required Qwen3-VL-32B intermediate-state encoder uses NVFP4/AWQ storage with FP16 computation; it is not replaced by a smaller model. The video VAE uses INT8 weights/BF16 computation and audio VAE remains FP32. Both compared engines use the same quantized checkpoints and Turbo adapter. Original full-precision quality parity is unproven.

The 15-second fox scene retains a consistent subject and continuous movement. The reviewer confirmed natural audio throughout. Short dialogue also passed listening review. Four-step pouring duplicates a bottle; eight-step pouring has excessive foam and incomplete placement. The pouring clips contain a brief native audio transient. These are retained limitations, not fixes applied to the soundtrack.

Hosted context processing, 2K regeneration, image/reference conditioning, larger canvases, other GPUs and unlisted durations/presets are outside this qualification. Unsupported kernel shapes fall back to native operations; loading is not proof of practical performance for those settings.

Downloads, artifacts and licenses

checkpoints.lock.json records immutable upstream revisions, sizes and SHA-256 values. Downloads preserve original bytes and reuse the persistent cache. No checkpoint conversion or calibration is required locally. ./run.sh download --verify-only --rehash verifies complete files. Optional reference weights under --profile all need substantially more disk.

The GHCR artifact image is ghcr.io/eliovp/paiton-vllm-plugin:minimax-h3-rdna4-v1.0.0. The private compiler is not required. Dockerfile reproduces that artifact image when the binary from the release bundle is present; Dockerfile.local prepares the complete ComfyUI engine locally. ./launch.sh --build rebuilds both from the bundle. The combined local image is not a distribution artifact.

MiniMax H3 uses a custom community license with territory and commercial-scale restrictions. Obtain the applicable rights before downloading or using its derivatives. Upstream model and encoder terms remain applicable; this runtime package does not relicense them. See THIRD_PARTY_NOTICES.md, retained licenses and MiniMax's license.