MiniMax M3 Promises 15.6x Faster AI Decoding

MiniMax, a prominent Chinese AI company, has unveiled technical details about its M2 series language models and teased innovations for its upcoming M3 series. The M3 models are set to dramatically enhance AI performance, offering up to a 15.6 times faster decoding speed at extremely long contexts of one million tokens. This advancement aims to make ultra-long-context AI agent deployment economically feasible by tackling the computational challenges of processing vast amounts of information.

The M2 series introduced an 'interleaved thinking' protocol, enabling models to alternate between planning and tool invocation while maintaining conversational context. This series was developed using 'Forge,' a scalable reinforcement learning system designed for agent-native training. Forge employs independent modules for agents, middleware, and training engines to manage complex, multi-step agent environments.

Furthering autonomous capabilities, the M2.7 model demonstrated remarkable self-evolution by managing 30% to 50% of its own development workflow. This included profiling training runs, diagnosing anomalies, and modifying its own codebase. On OpenAI's MLE Bench Lite suite, M2.7 achieved a 66.6% medal rate, performing comparably to Google's Gemini 3.1 Pro in autonomous ML research capabilities.