dgrauet/matrix-game-4.0-mlx

MLX format conversion of Skywork/Matrix-Game-4.0.

Converted with mlx-forge.

Measured with Matrix-Game-mlx on an M2 Pro (32 GB): quality preset ~27–38 s per 4-frame chunk at 24.4 GB peak; realtime preset ~3.5 s per chunk at 15 GB peak. The T5 encoder, Wan2.2 VAE and tokenizer are not in this repo: they are fetched from dgrauet/matrix-game-3.0-mlx at the revision pinned in split_model.json and checked by sha256.

Usage

These weights can be used with Matrix-Game-mlx:

git clone https://github.com/dgrauet/Matrix-Game-mlx
cd Matrix-Game-mlx/Matrix-Game-4-mlx
pip install -e .
# scenes: ancient-egypt | castle | racing — weights download automatically on first run
# quality preset (default; Wan VAE, 2 steps)
python -m matrix_game_mlx.inference.offline --scene ancient-egypt --input clip.mp4 --output output/egypt
# realtime preset (TAEHV codec, 1 step)
python -m matrix_game_mlx.inference.offline --scene castle --preset realtime --input clip.mp4 --output output/castle-realtime

Related Projects

• Upstream code: https://github.com/SkyworkAI/Matrix-Game
• Shared T5 encoder, Wan2.2 VAE and tokenizer (pinned): https://huggingface.co/dgrauet/matrix-game-3.0-mlx/tree/ff0d4df16dde5c86fca8cbd934a41fdff5eec00c

Files

• LICENSE (11.09 KB)
• config.json (1.11 KB)
• dit_ancient_egypt.safetensors (9.32 GB)
• dit_castle.safetensors (9.32 GB)
• dit_racing.safetensors (9.32 GB)
• driving_game/inference.json (894.00 B)
• driving_game/sink.png (1.17 MB)
• fps-castle/inference.json (807.00 B)
• fps-castle/sink.png (1.67 MB)
• fps-egypt/inference.json (828.00 B)
• fps-egypt/sink.png (1.42 MB)
• licenses/TAEHV-LICENSE (1.05 KB)
• split_model.json (1.89 KB)
• taehv.safetensors (21.79 MB)