Microsoft has released its MAI-Thinking-1 model, a generalist and reasoning model trained from scratch without third-party distillation. The model achieved high scores on benchmarks like AIME 2025 and SWE-Bench Pro, and demonstrated strong performance in human preference tests against existing models. Microsoft also detailed its "Frontier Tuning" approach for adapting models to specific workflows, claiming significant efficiency gains. AI
IMPACT Sets new SOTA on reasoning and coding benchmarks, while Frontier Tuning offers a path to efficient, customized enterprise models.
RANK_REASON Microsoft released a new frontier model, MAI-Thinking-1, with detailed technical information and benchmark results. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- AIME 2025
- Sonnet 4.6
- Frontier Tuning
- Gemma 4 12B
- GPT-5.4
- GPT-Image-2
- Ideogram 4.0
- MAI-Code-1-Flash
- MAI-Image-2.5
- MAI-Thinking-1
- Microsoft
- SWE-Bench Pro
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
