EigenLabs/Qwen3.5-9B-MLX-4bit-mtp

Qwen3.5-9B MLX W4/g64 + Embedded W4 MTP

This is a full vision-language MLX conversion of Qwen/Qwen3.5-9B with the model's native one-layer MTP proposal head embedded in the same indexed safetensors weight set under the mtp.* prefix.

• Language target and vision tower: byte-identical to mlx-community/Qwen3.5-9B-MLX-4bit (W4/g64 affine, BF16 scales/biases).
• MTP head: the 15 native mtp.* tensors from the source checkpoint, RMSNorm weights converted to the MLX x·w convention (+1) and the 8 linear modules quantized to W4/g64 affine with BF16 scales/biases (~137 MB).
• Declared via mtplx_mtp / mtplx_mtp_quantization in config.json (prefix: "mtp.", block_size: 3). The head shares the target embedding and LM head, as intended by the source architecture.

There is no separate MTP repository: the proposal head travels inside the checkpoint, so the draft/target pairing cannot drift.

Measured (M4 Max, 546 GB/s, temperature 0)

Configuration	decode tok/s
target only	47.9–48.5
embedded MTP	73–83

Conversion notes

Raw source mtp.* norms are stored in the Qwen3-Next x·(1+w) convention and MUST be shifted by +1 for MLX runtimes; quantization scales/biases MUST be bfloat16. Raw norms produce a head whose drafts are rejected (slower than no MTP at all); float16 scales fail the assistant load, which serving stacks should surface as a logged fallback to target-only decoding.