How to Run MiniMax‑M2.7 on Mac: Comparing Two Quantization Paths

Overview

MiniMax-M2.7 is a 228.7 B‑parameter Mixture‑of‑Experts (MoE) language model with a 192K context window. Each token activates roughly 10 B parameters. Reported benchmark scores include SWE‑Pro 56.22 % and MLE Bench Lite 66.6 %.

Why standard MLX uniform quantization fails

Uniform quantization of the entire model in the MLX ecosystem reduces MMLU accuracy to about 25 % because the router gate is also quantized, causing tokens to be routed to incorrect experts.

Path 1 – JANGTQ + MLX Studio (recommended)

JANGTQ (JANG TurboQuant) is a mixed‑precision quantization scheme that keeps the router gate, attention layers, and shared experts at 8‑bit or fp16 while compressing the expert MLP (≈98 % of parameters) with a 2‑bit codebook and Hadamard rotation.

Installation and inference example

pip install jang-tools

from huggingface_hub import snapshot_download
from jang_tools.load_jangtq import load_jangtq_model
from mlx_lm import generate

model_path = snapshot_download("JANGQ-AI/MiniMax-M2.7-JANGTQ")
model, tokenizer = load_jangtq_model(model_path)

messages = [{"role":"user","content":"用5句话解释光合作用"}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
out = generate(model, tokenizer, prompt, max_tokens=600, verbose=True)

# Strip reasoning chain
if "</think>" in out:
    out = out.split("</think>")[-1].strip()
print(out)

Hardware requirements and performance

Minimum RAM: 64 GB (96 GB recommended) on Apple Silicon.

Disk footprint: 56.5 GB.

MMLU (200‑question) ≈ 91.5 %.

Speed on M3 Ultra: ~44 tokens/s.

Performance per Apple Silicon model:

M3 Ultra / M2 Ultra (96 GB RAM) – ~44 tok/s.

M4 Max (96 GB RAM) – ~35‑40 tok/s.

M4 Pro (64 GB RAM) – ~25‑30 tok/s (tight).

Path 2 – LM Studio + GGUF (simpler)

LM Studio ships a pre‑quantized GGUF version of MiniMax‑M2.7 built with llama.cpp b8778. The GGUF file is available from lmstudio-community/MiniMax-M2.7-GGUF.

Default generation parameters:

Temperature = 1.0 (required).

Top K = 40.

Top P = 0.95.

Steps

Download and install LM Studio from https://lmstudio.ai/download.

Search for minimax/minimax-m2.7 and select the GGUF version.

Set the parameters above.

Start a chat.

LM Studio reports a minimum system memory requirement of 138 GB. On a 96 GB Mac the model can run but MMLU drops to roughly 64‑65 % for the 4‑bit version.

Comparison

Disk usage : JANGTQ 56.5 GB vs GGUF ≈ 108 GB.

Minimum RAM : JANGTQ 64 GB vs GGUF 138 GB.

MMLU quality : JANGTQ 91.5 % vs GGUF ~64‑65 % (4‑bit).

Speed on M3 Ultra : JANGTQ ~44 tok/s; GGUF not yet measured.

Ease of use : JANGTQ requires installing jang-tools; GGUF works out‑of‑the‑box.

Ecosystem compatibility : JANGTQ integrates with the MLX ecosystem; GGUF provides an OpenAI‑compatible API.

Key settings reminder

Temperature must be set to 1.0 – a temperature of 0 causes the always‑reasoning chain to loop indefinitely inside <think> tags.

max_tokens ≥ 8192 – the always‑reasoning mode needs sufficient token budget.

System RAM must exceed the model file size – otherwise the model swaps to disk, causing a drastic speed drop.

Conclusion

For local deployment of MiniMax‑M2.7 on Apple Silicon, JANGTQ + MLX Studio offers the smallest footprint (56.5 GB) and the highest quality (2‑bit quantization achieving 91.5 % MMLU). LM Studio provides a more user‑friendly, out‑of‑the‑box experience but requires substantially more memory and yields lower accuracy.