Qwen3.6-27B-Q4-MTPLX

An Apple Silicon MLX checkpoint derived from Qwen/Qwen3.6-27B. It includes an MTPLX sidecar for native speculative decoding.

This revision uses an optimized mixed-low-bit trunk. On the local WikiText-2 128 x 512 evaluation its perplexity was 8.3781. The complete repository is approximately 14.57 GB.

Standard MLX usage

from mlx_lm import load, generate

model, tokenizer = load("wang-yang/Qwen3.6-27B-Q4-MTPLX")
print(generate(model, tokenizer, prompt="The capital of France is", max_tokens=32))

For speculative decoding, use MTPLX and run mtplx tune on the target Mac. Observed MTP acceptance and speed are workload-dependent, so applications should also benchmark representative prompts.