MiniMax Unveils M3 AI Model With 15x Speed Boost and Self-Developing Agent Capabilities

MiniMax unveils its M3 AI model series featuring a groundbreaking 15.6x decoding speed boost and a self-developing agent that autonomously managed 30-50% of its own development workflow, rivaling Google's Gemini 3.1 Pro on key benchmarks.

• MiniMax is teasing its upcoming M3 model series, which introduces a new MiniMax Sparse Attention (MSA) mechanism that delivers a 15.6x decoding speed boost and 9.7x prefilling speedup at one-million token context lengths, making ultra-long-context AI agent deployment economically viable.
• A newly released technical report on the M2 model series reveals key engineering innovations, including a 229.9 billion parameter sparse Mixture-of-Experts architecture that activates only 9.8 billion parameters per token, and an 'interleaved thinking' protocol that enables autonomous agents to plan, use tools, and recover from errors within a single workflow.
• MiniMax built a reinforcement learning system called 'Forge' to train long-horizon agent behavior, featuring optimizations like Prefix Tree Merging for up to 40x training speedup, ultimately producing the M2.7 checkpoint, which autonomously handled 30-50% of its own development workflow and achieved a 66.6% medal rate on OpenAI's MLE Bench Lite, tying Google's Gemini 3.1 Pro.