SceneWorks/mochi-1-mlx

Mochi 1 — MLX (SceneWorks)

Apple-Silicon (MLX) repackage of genmo/mochi-1-preview for the SceneWorks native MLX runtime (mlx-gen-mochi). The AsymmDiT transformer ships as three self-contained, pre-quantized tiers; the T5-XXL text encoder and AsymmVAE are shared across tiers (hosted once as siblings, not duplicated per tier).

Layout

q4/          transformer/model.safetensors  + split_model.json + quantize_config.json   (~9 GB)
q8/          transformer/model.safetensors  + split_model.json + quantize_config.json   (~12 GB)
bf16/        transformer/model.safetensors  + split_model.json                          (~19 GB)
text_encoder/  T5-XXL, bf16 (2-shard) + model.safetensors.index.json                    (~9.5 GB)
vae/           AsymmVAE decoder, bf16                                                    (~0.9 GB)
tokenizer/     T5 SentencePiece tokenizer

To load a tier, point the loader at the tier dir (e.g. q4/); the shared text_encoder/, vae/, and tokenizer/ resolve from the parent as siblings of the tier subdirs.

Tiers

Tier	Transformer precision	Approx. size
q4	4-bit affine (group 64), 573 quantized Linears	~9 GB
q8	8-bit affine (group 64), 573 quantized Linears	~12 GB
bf16	dense bf16	~19 GB

Each tier's split_model.json records the quant selection (quantized, quantization_bits, quantization_group_size); a tier dir is the quant selection (no on-the-fly requantization).

Shared components

• text_encoder/ — T5-XXL (google/t5-v1_1-xxl, 24-layer encoder, d_model=4096) hosted in bf16 (2-shard set + a matching model.safetensors.index.json), bit-identical to round-to-bf16 of the upstream fp32 weights (the precision the reference pipeline loads at).
• vae/ — AutoencoderKLMochi decoder in bf16.
• tokenizer/ — the T5 SentencePiece tokenizer.

Source & license

Repackaged from genmo/mochi-1-preview. Licensed Apache-2.0, matching the upstream Mochi 1 release. Weights are unchanged apart from the per-tier quantization of the transformer and the dtype/shard repackaging described above.