Hasty Briefsbeta

Bilingual

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.