MAI-Thinking-1
4 hours ago
- #enterprise AI
- #AI reasoning model
- #machine learning development
- Microsoft AI introduces MAI-Thinking-1, a medium-sized reasoning model with 35B active and ~1T total sparse Mixture of Experts parameters.
- The model matches leading models on software engineering benchmarks like SWE-Bench Pro, shows advanced math reasoning (e.g., 97.0% on AIME 2025), and is preferred over Claude Sonnet 4.6 in blind human evaluations.
- Trained from scratch using clean, commercially licensed data without distillation from third-party models, emphasizing learning over imitation.
- Part of the Hill-Climbing Machine philosophy: capabilities improve through co-designed pipeline with clean data, self-sufficient stack, and learned intelligence.
- Enterprise-ready with 256k token context window, function calling, Chat Completions API compatibility, and security through Microsoft Foundry.
- Balances safety and helpfulness via reinforcement learning, treating unsafe compliance as defects, without unnecessary refusals.
- Aims for Humanist Superintelligence—AI serving people and organizations, not replacing them, while upholding human autonomy.