tokenfires/Qwen3.6-35B-A3B-MLX-4bit

4-bit MLX quantization of Qwen/Qwen3.6-35B-A3B for Apple Silicon. Quantization: affine, 4 bits, group size 64. The MoE router gates and shared-expert gates are kept at 8 bits to preserve routing accuracy.

Use with LM Studio

Search for tokenfires/Qwen3.6-35B-A3B-MLX-4bit in the LM Studio model downloader, or open this page and choose Use this model → LM Studio.

Use with mlx-lm

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("tokenfires/Qwen3.6-35B-A3B-MLX-4bit")

messages = [{"role": "user", "content": "hello"}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, verbose=True)