# Start server at http://localhost:8080/v1/chat/completions
uvx --from mlx-vlm --with torchvision \
  mlx_vlm.server \
  --host 127.0.0.1 \
  --port 8080 \
  --model spicyneuron/Qwen3.6-35B-A3B-MLX-4.8bit-vision

Benchmarks

metric	mlx-community/ Qwen3.6-35B-A3B-4bit	mlx-community/ Qwen3.6-35B-A3B-4.4bit-msq	4.8 bit (this model)	5.4 bit
bpw	4.503	4.787	4.788	5.438
peak memory (1024/512)	20.683	21.922	21.928	24.741
prompt tok/s (1024)	2719.4470 ± 15.2250	2695.9370 ± 12.5260	2734.5260 ± 3.8810	2665.3060 ± 11.4520
gen tok/s (512)	108.4990 ± 0.4910	94.2940 ± 0.3650	97.2820 ± 0.0800	89.4920 ± 0.2610
kl divergence	0.0838 ± 0.0008	0.1689 ± 0.0015	0.0244 ± 0.0004	0.0189 ± 0.0003
perplexity	4.6150 ± 0.0320	4.2490 ± 0.0280	4.6410 ± 0.0320	4.6440 ± 0.0320
hellaswag	0.5560 ± 0.0220	0.5780 ± 0.0220	0.5440 ± 0.0220	0.5370 ± 0.0110
piqa	0.7940 ± 0.0180	0.7920 ± 0.0180	0.7920 ± 0.0180	0.7980 ± 0.0180
winogrande	0.7260 ± 0.0200	0.7400 ± 0.0200	0.7120 ± 0.0200	0.7100 ± 0.0200

I've moved over to using speed + KL divergence as my primary optimization metrics. Hellaswag, PIQA, Winogrande, and perplexity are kept as sanity checks, though these require high sample sizes to get usable signal.

Tested on a Mac Studio M3 Ultra with:

mlx_lm.convert --hf-path Qwen/Qwen3.6-35B-A3B --mlx-path ./mlx && mlx_lm.kld --baseline-model ./mlx
mlx_lm.perplexity --sequence-length 512 --seed 123
mlx_lm.benchmark --prompt-tokens 1024 --generation-tokens 512 --num-trials 5
mlx_lm.evaluate --tasks hellaswag --seed 123 --num-shots 0 --limit 500
mlx_lm.evaluate --tasks piqa --seed 123 --num-shots 0 --limit 500
mlx_lm.evaluate --tasks winogrande --seed 123 --num-shots 0 --limit 500

Required PRs:

Methodology

Quantized with a mlx-lm fork, drawing inspiration from Unsloth/AesSedai/ubergarm style mixed-precision GGUFs. MLX quantization options differ than llama.cpp, but the principles are the same:

Sensitive layers like MoE routing, attention, and output embeddings get higher precision
More tolerant layers like MoE experts get lower precision