spicyneuron/Qwen3-Next-Coder-MLX-mixed-4.5-bit

Qwen3-Coder-Next optimized for MLX. Note: Uses MXFP4 for some module paths.

EDIT: v2 fixes some misassigned shared expert gates. Slower, but with 4x better perplexity.

EDIT: v3 bumps edge experts to Q8 for further perplexity improvement and minimal effect on speed.

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-Next-Coder-MLX-mixed-4.5-bit

Methodology

Quantized using a custom script inspired by 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

This one is comparable to Unsloth's UD-Q4_K_XL Unsloth's MOE-MXFP4 in size, but loads and runs noticeably faster thanks to MLX.

Benchmarks

• unsloth/Qwen3-Coder-Next-GGUF:UD-Q4_K_XL
• mlx-community/Qwen3-Coder-Next-4bit
• Qwen3-Next-Coder-MLX-mixed-4.5-bit (v1)
• Qwen3-Next-Coder-MLX-mixed-4.5-bit (v2, ~4.4 bit)
• Qwen3-Next-Coder-MLX-mixed-4.5-bit (v3, ~4.9 bit)

Prompt Processing (tokens/sec)

Prompt Size	GGUF	MLX 4bit	MLX 4.5bit (v1)	MLX 4.4bit (v2)	MLX 4.9bit (v3)
1000	1440.60	1917.29	1894.38	1871.55	1868.77
5000	1511.29	2113.98	2069.36	2079.87	2071.76
10000	1491.41	2073.89	2032.13	2039.11	2031.04
20000	1387.15	1888.56	1854.83	1860.35	1854.24

Generation (tokens/sec)

Gen Size	GGUF	MLX 4bit	MLX 4.5b (v1)	MLX 4.4b (v2)	MLX 4.9b (v3)
500	49.35	76.39	75.30	66.82	67.19
1000	49.12	74.67	73.16	65.86	64.82
2000	49.01	71.99	70.95	63.68	62.82
5000	48.64	67.72	66.67	61.04	60.99

Perplexity (MLX Quants)

Model	Perplexity	Relative	Relative %
MLX 4bit	4.118 ± 0.021	—	—
MLX 4.5bit (v1)	4.096 ± 0.021	-0.022	-0.53%
MLX 4.4bit (v2)	4.024 ± 0.021	-0.094	-2.28%
MLX 4.9bit (v3)	4.016 ± 0.021	-0.102	-2.48%

# llama.cpp 8130
llama-bench -fa 1 --batch-size 2048 --ubatch-size 2048 --repetitions 5

# mlx_lm v0.30.7
mlx_lm.benchmark --num-trials 5
mlx_lm.perplexity --sequence-length 1000 --seed 222