from mlx_lm import load, generate
model, tokenizer = load("bearzi/gemma-4-26B-A4B-it-oQ8")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Hello"}],
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=True))

About oQ

Quantization: oQ8 (sensitivity-driven mixed precision, group_size=64)
Format: MLX safetensors
Compatible with: mlx-lm, mlx-vlm, oMLX on Apple Silicon

oQ measures per-layer quantization sensitivity through calibration and allocates bits where they matter most — critical layers stay at higher precision, tolerant layers compress aggressively. Target averages of 2/3/4/6/8 bits are provided; actual per-layer bits vary by measured sensitivity.

See oQ documentation.

Comparative benchmarks and feedback welcome — please open a discussion.