GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Pilcothink
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Pilcothink
Model Tree
MiniMaxAI/MiniMax-M3
Input Modalities
Output Modalities
Supported Functionality
| Item | Value |
|---|---|
| Base model | MiniMaxAI/MiniMax-M3 |
| Quantized repo | Pilcothink/MiniMax-M3-int4-AutoRound |
| Local source folder | MiniMax-M3-W4A16-RTN |
| Quantization method | AutoRound |
| Quantization format | int4 / W4A16 |
| Weight precision | 4-bit |
| Activation precision | 16-bit |
| Base model relation | quantized |
| License | minimax-community, inherited from upstream model |
The model card has been uploaded first.
The quantized model weight files are still being uploaded and may not be available yet.
This repository is intended to host int4 AutoRound quantized weights for MiniMax-M3.
Benchmark numbers, runtime requirements, local deployment notes, and usage examples in the upstream section below describe the original MiniMaxAI/MiniMax-M3 model unless explicitly stated otherwise.
This model is a quantized derivative of MiniMaxAI/MiniMax-M3, which is released under the minimax-community license.
The original model card is preserved below for attribution and reference.
============================================================
============================================================
The following section is copied from the upstream model card of
MiniMaxAI/MiniMax-M3.
It describes the original model, not this int4 AutoRound quantized release.
MiniMax-M3 is a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.
Highlights:
M3 is powered by MiniMax Sparse Attention (MSA), a high-performance sparse attention operator designed for million-token contexts. Compared with GQA, MSA dramatically reduces the attention compute and memory footprint while preserving model quality.
📄 Read the technical report: arXiv:2606.13392 · Hugging Face Papers
M3 supports three reasoning modes through the thinking parameter:
enabled — Reasoning is always enabled.adaptive — M3 automatically determines when additional reasoning is beneficial.disabled — Reasoning is disabled to minimize latency and maximize throughput.Download the model:
bash
hf download MiniMaxAI/MiniMax-M3 --local-dir MiniMax-M3
We recommend the following inference frameworks (listed alphabetically) to serve the model:
SGLang - see SGLang cookbook.
vLLM - see vLLM recipes.
Transformers - see Transformers docs.
We recommend the following parameters for best performance: temperature=1.0, top_p=0.95, top_k=40.
Contact us at model@minimax.io.