Modelpublic

ddalcu/MiMo-V2.6-Distill-Qwen-9B-MLX-Serve-4bit

sourceHugging Faceupdated 5d agoView on Hugging Face

0likes162downloads

Model Card

MiMo-V2.6-Distill-Qwen-9B MLX 4-bit

4-bit affine (group size 64) MLX conversion of XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B for mlx-serve. 7.4 GB.

—Arch is Qwen3.5-9B (qwen3_5), so it runs on mlx-serve with no special flags.
—Vision tower and token embeddings are kept in bf16. Image input works.
—Thinking, tool calling and streaming tested.
—MTP head grafted from the base Qwen/Qwen3.5-9B (the upstream checkpoint ships none). mlx-serve uses it for speculative decoding, 2.3x on predictable text in llmprobe.

bash

mlx-serve pull ddalcu/MiMo-V2.6-Distill-Qwen-9B-MLX-Serve-4bit
mlx-serve run ddalcu/MiMo-V2.6-Distill-Qwen-9B-MLX-Serve-4bit

License and usage terms follow the upstream model.