BobClawblaw/MiniMax-M3.1-Flash-MLX-mix34

MiniMax-M3.1-Flash, mixed 4/3-bit (mix34)

MiniMaxAI/MiniMax-M3.1-Flash (revision 54e4e69d) as a community checkpoint for TensorFold's native Metal engine (its minimax_m3 family): a lossy conversion of the release, sized to serve on one Apple M5 Ultra with 256 GB. Its window on 256 GB is 32,768 tokens; mix256k trades quality for a longer one.

Serving: TensorFold's minimax_m3 family (pull request #1235, under review) with the recipe BobClawblaw/minimax-m31-flash-mlx-m5-ultra. config.json is the release's own (model_type minimax_m3_vl, the language model under text_config), with the MLX quantization block in place of the release's and the drafters' dflash2 and dspark blocks beside it. The checkpoint does not load in mlx-lm or mlx-vlm: the layout is MLX affine quantization, but the model is TensorFold's.

What is in it

• Routed experts: 4-bit affine, groups of 128, on the 26 layers that gain most (19-31, 33, 40, 42-44, 48-50, 52, 53, 57-59 of 60, ranked by each layer's output error on captured activations); 3-bit affine, groups of 64, on the rest. Each width is a least-squares refit of the release's NVFP4 experts (scales and biases fitted to the NVFP4 values, not a plain round-to-nearest).
• Attention (q, k, v and the lightning indexer's projections), shared expert, dense layers, embeddings and the output head: 8-bit affine, groups of 64, from the release's MXFP8 (exact in bf16 before quantizing).
• Router: fp32, as released.
• Vision tower and projector (model-vision.safetensors): bf16 as released (the patch convolution fp32).
• Drafters: the release's own DSpark head (model-dspark.safetensors) and incoai/MiniMax-M3.1-Flash-DFlash2 (revision ac07a7c1, Apache-2.0, model-dflash2.safetensors), its projections in the same affine layout.
• About 204 GiB in all (DSpark and DFlash2 included; the engine loads one drafter).

Quality

Next-token KL divergence and top-1 agreement against the release's own NVFP4 weights, over the 1,024 positions of one chat, teacher-forced, with TensorFold's exact bf16 key/value cache (engine cb35b7c28):

Checkpoint	Mean KL	Top-1 agreement
mix34	0.244	80.7%
mix256k	0.295	78.4%

The recipe reports the quality suites and speeds.

License

The MiniMax Community License: free for non-commercial use; commercial use needs the attribution and the notice or authorization it describes. The DFlash2 drafter's weights are incoai's, under Apache-2.0.