Researchers have introduced the MiniMax-M2 series, a new family of Mixture-of-Experts language models designed for agentic deployment. The flagship M2 model boasts 229.9 billion total parameters but activates only 9.8 billion per token, emphasizing efficiency. This series is built on agent-driven data pipelines, a scalable agent-native reinforcement learning system called Forge, and a checkpoint (M2.7) that demonstrates early self-evolution capabilities by debugging training runs. The MiniMax-M2 series achieves frontier-tier performance on various agentic benchmarks, including coding, deep search, and office tasks. AI
IMPACT Introduces a new model architecture focused on efficiency and agentic capabilities, potentially influencing future LLM development for specialized tasks.
RANK_REASON The cluster describes a new research paper detailing a novel series of language models.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
