spicyneuron/Gemma-4-26B-A4B-MLX-4.7bit-vision

Gemma-4-26B-A4B-it optimized for MLX. This quant supports image input and requires a vision-enabled MLX server.

EDIT April 5, 2026: Updated model for better performance / speed tradeoff, plus full precision vision embedding. Added benchmarks.

Usage

# Start server at http://localhost:8080/chat/completions
uvx --from mlx-vlm --with torchvision \
  mlx_vlm.server \
  --host 127.0.0.1 \
  --port 8080 \
  --model spicyneuron/Gemma-4-26B-A4B-MLX-4.7bit-vision

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

Benchmarks

metric	mlx-community/gemma-4-26b-a4b-it-4bit	unsloth/gemma-4-26b-a4b-it-UD-MLX-4bit	4.7 bit (this model)
bpw	4.587	4.743	4.704
peak mem	15.312	15.804	15.681
prompt proc (1024)	2514.427	2508.474	2501.617
token gen (512)	97.214	92.890	91.860
perplexity	146.663 ± 0.921	211.070 ± 1.378	133.539 ± 0.827
hellaswag	0.525 ± 0.011	0.531 ± 0.011	0.532 ± 0.011
piqa	0.72 ± 0.01	0.712 ± 0.011	0.719 ± 0.01
winogrande	0.635 ± 0.014	0.635 ± 0.014	0.639 ± 0.014

• Bits per weight calculated against only the language_model weights.
• Perplexity in Gemma 4 was surprisingly high but seemed consistent across my trials. Could be a side effect of using allenai/tulu-3-sft-mixture. Best to interpret it as weaker signal than the other benchmark results.

Tested with:

mlx_lm.perplexity --sequence-length 4096 --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 2000
mlx_lm.evaluate --tasks piqa --seed 123 --num-shots 0 --limit 2000
mlx_lm.evaluate --tasks winogrande --seed 123 --num-shots 0 --limit 2000