MiniMax M3

MiniMax

Open-weight model · runs on-premises · Published on 5 July 2026 · Updated on 3 September 2026

MiniMax M3 handles long context and multimodal input at low cost. It totals 428 billion parameters, of which 23 billion are active, with a one-million-token context window and sparse attention. It suits large documents and image content.

Key points: Multimodal · 428B mixture-of-experts · 1M context · Sparse attention · Cost-efficient.

Architecture

MiniMax M3 totals 428 billion parameters, of which 23 billion are active (128 experts, 4 active per token), 60 layers, a 1,048,576-token context and 854 GB of weights published in BF16; natively multimodal (text, image, video), it relies on MiniMax Sparse Attention, which lowers the cost of long sequences. MiniMax Community licence: commercial use requires the “Built with MiniMax M3” notice and written authorisation above 20 million dollars of yearly revenue (Hugging Face card and config.json re-read on 2 September 2026).

Strengths

It handles long context and multimodal input at low cost, covering both text and images.

Use cases

MiniMax M3 suits large documents, image analysis, and workflows where cost per token is the priority.

A public deployment: mmr.qdna.fr alternates MiniMax M3 with GLM 5.3 to arbitrate its 227 analyses, and publishes the majority vote of the three draws.

Deployment

Served by vLLM on a GPU server, or consumed via API. It integrates with the platform through LiteLLM.

Every model on the platform is interchangeable through a single gateway: switching models is a matter of changing one configuration line, with no code to rewrite. Local execution remains the priority.

Official sources

• MiniMaxAI/MiniMax-M3, official model card on Hugging Face: licence, parameters, context window and weight format
• Artificial Analysis, model comparison index