MiniMax-M3
#2 in Small & Edge Modelsminimax · since 2026-06-01 · 52× · last seen Jul 01, 2026
95
Momentum
MiniMax-M3 is an open-weight multimodal language model released by MiniMax on June 1, 2026. It is built on a Mixture-of-Experts architecture with approximately 428 billion total parameters and approximately 23 billion active parameters per forward pass. Its core architectural innovation is MiniMax Sparse Attention (MSA), which replaces standard quadratic attention to enable a 1-million-token context window at drastically reduced compute cost. M3 is the first open-weight model to simultaneously deliver frontier-level coding, a 1M-token context window, and native multimodality (text, image, video).
Momentum trend
04.04.03.07.
Features
| Context Window (Tokens) | Up to 1,048,576 tokens (1M); guaranteed minimum 512,000 tokens. Output limit: up to 512,000 tokens. Inputs >512K tokens billed at a higher rate. |
| License | MiniMax Community License (not standard open-source like Apache 2.0 or MIT; commercial use requires separate review of license terms) |
| Platform | MiniMax API (platform.minimax.io), MiniMax Code (agent product), OpenRouter, ModelScope; self-hosting via SGLang, vLLM, Transformers, KTransformers (Hugging Face: MiniMaxAI/MiniMax-M3) |
| Price | Pay-as-you-go: from $0.30/M input tokens. Token plan subscriptions: Plus $20/month (~1.7B tokens), Max $50/month (~5.1B tokens), Ultra $120/month (~9.8B tokens). |
| Release Date | June 1, 2026 (API launch); weights on Hugging Face from June 7, 2026 |