MiniMax-M2.7-oQ8e

Even as of July 2026 on M3 Mac 512Gb RAM I can't seem to find something better than Minimax 2.7, so here's the OMLX 8-bit oQ8e optimized quant, on paper it should be better than a regular 8-bit quant available elsewhere on HF.

The optimized 8-bit quant runs quite fast and with turboquant you can pretty easily fit other models on a 512Gb or at least get a huge KV cache size while gaining up to 20 TPS vs 15 on the unquantized model; the difference between annoying and quite practical.

Below is the original base model card...

mlx-community/MiniMax-M2.7

This model mlx-community/MiniMax-M2.7 was converted to MLX format from MiniMaxAI/MiniMax-M2.7 using mlx-lm version 0.31.3.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/MiniMax-M2.7")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, return_dict=False,
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)