spicyneuron/Qwen3.5-397B-A17B-MLX-2.6bit

Qwen3.5-397B-A17B optimized for MLX!

• Mixed-precision quantization balances throughput, accuracy, and memory.
• Similar quality to a 4-bit baseline but requires 40% less memory.
• Fixed chat template allows more reliable prompt caching.
• This version does NOT support vision (image input).

Also available as a larger 178GB version: https://huggingface.co/spicyneuron/Qwen3.5-397B-A17B-MLX-3.5bit

Usage

# Start server at http://localhost:8080/v1/chat/completions
uvx --from mlx-lm mlx_lm.server \
  --host 127.0.0.1 \
  --port 8080 \
  --model spicyneuron/Qwen3.5-397B-A17B-MLX-2.6bit

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.

Benchmarks

metric	lmstudio-community 4 bit	2.6 bit	3.5 bit
perplexity	3.919 ± 0.019	3.852 ± 0.018	3.919 ± 0.019
hellaswag	0.594 ± 0.022	0.598 ± 0.022	0.622 ± 0.022
piqa	0.798 ± 0.018	0.802 ± 0.018	0.804 ± 0.018
winogrande	0.744 ± 0.02	0.718 ± 0.02	0.746 ± 0.019
p1024/g512 prompt	490.702	489.545	479.453
p1024/g512 gen	39.192	38.398	35.547
p1024/g512 mem	225.095	131.523	179.842