Gemma 4 26B-A4B on a 64 GB Mac: speed, quality, memory
Google, released Apr 2026. 25.2B parameters, 3.8B active (mixture of experts). 4-bit MLX build, 15.4 GB download. License Apache-2.0.
- HumanEval+ pass@1: 149 of 164 (90.9%), rank 3 of 8
- HumanEval pass@1: 157 of 164 (95.7%)
- Generation speed: 80.7 tokens per second (5 trials, 80.6 to 80.7); prompt 1,562 tokens per second
- Peak memory: 15.1 GB · power while generating: 27 W · GPU peak 79 °C
Run it yourself
pip install mlx-lm mlx_lm.chat --model mlx-community/gemma-4-26b-a4b-it-4bit
HumanEval/73
Count the fewest changes needed to make a list read the same both ways.
It left a broken first draft in the answer. 'n - - 1' means n plus one, so the code crashed before its corrected loop ran.
Problems it failed
HumanEval/73, HumanEval/103, HumanEval/124, HumanEval/129, HumanEval/132, HumanEval/145, HumanEval/156
Raw files: speed.log, evaluate.log, eval_results.json