shadow/mmh3turbo-bundles

mmh3turbo bundles

Prepared weights for mmh3turbo, an MLX port of MiniMax-H3 that generates video and audio on Apple Silicon using hand-written int8 Metal kernels.

You do not need to download these by hand:

uvx mmh3turbo "a rainbow colored skunk leaps over a mossy log in a supermarket"

Contents

File	Size	What it is
dit.bin + dit.idx	19.5 GiB	The 50-block DiT as per-output-channel int8, requantized from the reference fp8-E4M3 checkpoint. Laid out for this port's fused Metal kernels, so it is not interchangeable with MLX-native quantized repos.
qwen3vl_4bit.safetensors	14.2 GiB	The Qwen3-VL-32B text tower's first 50 layers in MLX 4-bit (group 64). Cuts text encoding from 43.8 s to 3.3 s.

Why a 4-bit text tower

The tower is consumed from a 7.9 GB GGUF rather than the 66.7 GB unquantized encoder, and streaming it costs ~52 s per new prompt because its 50 layers are 25.16 B elements and numpy k-quant dequantization runs at ~487 M/s. Quantizing once to MLX 4-bit moves the matmuls onto the GPU and turns that into an mmap'd load. The source is Q2_K — 2.6 effective bits — so 4 bits loses nothing against it: cosine similarity against the streamed path is 0.9994.

What the DiT bundle is not

This is int8 with per-output-channel scales plus per-token activation scales, which is what gemm_i8 and flash_i8 consume. It will not load in implementations built on mx.quantize. SmoothQuant was tried and rejected: it made the end-to-end velocity error worse (16.75% -> 63.43%), because the source is already fp8-E4M3 with 3 mantissa bits and has no range left to migrate.

License

MIT for the packaging. The underlying weights are MiniMax-H3's; see MiniMaxAI/MiniMax-H3.