A mixed-precision quant that balances speed, memory, and accuracy.
4-bit baseline with important layers at higher precision.
This quant supports image input and requires a vision-capable server.

Usage

# Start server at http://localhost:8080/v1/chat/completions
uvx --from mlx-vlm mlx_vlm.server \ 
  --host 127.0.0.1 \
  --port 8080 \
  --model spicyneuron/Qwen3.5-4B-MLX-6.8bit-vision

Benchmarks

metric	mlx-community/Qwen3.5-4B-MLX-4bit	mlx-community/Qwen3.5-4B-OptiQ-4bit	6.8 bit (this model)
bpw	4.503	6.219	6.750
base memory	2.205	3.045	3.305
peak memory (1024/512)	3.625	4.541	4.814
prompt tok/s (1024)	2674.115 ± 3.462	2661.445 ± 1.504	2589.364 ± 13.109
gen tok/s (512)	166.004 ± 0.312	135.779 ± 0.031	127.137 ± 0.121
kl mean	0.083 ± 0.004	0.041 ± 0.002	0.007 ± 0.001
kl p95	0.180 ± 0.004	0.100 ± 0.002	0.020 ± 0.001
perplexity	5.053 ± 0.035	4.932 ± 0.034	4.921 ± 0.034

Tested on a Mac Studio M3 Ultra with:

mlx_lm.convert --hf-path ... --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

Required PRs:

Methodology

Quantized with a mlx-vlm fork. 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