At Build 2026 in June 2026, Microsoft AI launched MAI-Thinking-1, their first self-developed reasoning model. With 35 billion active parameters and a 128K context window, this model was trained from scratch—no distillation from other models—using commercially licensed data throughout.
Clear Objectives
Microsoft targets MAI-Thinking-1 at multi-step agentic tasks such as tool use, function calling, and reasoning workflows, which are important features for future AI agent development. The design of this model shows that Microsoft envisions something beyond just ordinary LLMs.
Impressive Benchmark Results
Testing results on SWE-Bench Pro (coding benchmark) found that MAI-Thinking-1 scored close to Anthropic's Claude Opus 4.6, which is an impressive achievement for a first-generation developed model. Additionally, Microsoft announced MAI-Code-1-Flash, a 5B coding model that performs comparably to Haiku but at a lower cost.
Different Strategy
Interestingly, Microsoft didn't use distillation from other models but chose to develop their models from scratch. This strategy helps avoid potential copyright issues. In this round, Microsoft announced 7 new self-developed models, showing their commitment to building their own ecosystem.
Entering the Reasoning Model Competition
The launch of MAI-Thinking-1 officially announces that Microsoft is ready to enter the reasoning model field currently dominated by OpenAI, Google, and Anthropic. That Microsoft, OpenAI's partner, is also developing its own models shows that the AI market is important enough that giants need their own solutions. Microsoft's entry into this field will make AI competition more interesting and could lead to faster technology development, as Microsoft has the resources and ability to invest heavily, which will push competitors to continuously develop their technology.
Reference:
https://card-bitcoin.com/ai-model-release-tracker-microsoft-ais-first-reasoning-model-arrives/