Microsoft’s AI strategy for a while has just been betting big on OpenAI, shipping its models through Copilot, and reaping the profits. Now, things have become more complicated as Microsoft is creating AI models itself.
Also read: Snowflake-Anthropic deepen AI handshake as India emerges as key market
On June 2, the in-house Microsoft AI lab introduced several models with the leading one being MAI-Thinking-1, an AI reasoning model aimed at directly competing with the state-of-the-art midweight model available today, and MAI-Code-1-Flash, a lightweight coding model that was immediately deployed into GitHub Copilot. All of these mark the arrival of an unprecedented development for Microsoft – credible AI stacks created in-house.
What distinguishes these new developments is not only the functionality of these new models themselves but also the way they were developed. In particular, both of these models were trained from scratch using clean data purchased from commercial providers without any distillation from other models, which is in direct contrast with industry norms when companies build their AI on existing models’ intelligence. The approach employed by Microsoft here has been dubbed “capabilities should be learned, not inherited,” and that approach has a clear rationale as models learn behavior better than they mimic others.
MAI-Thinking-1 is the main event. This is an extremely sparsely populated Mixture of Experts model with 35 billion active parameters, which is much smaller in the size of its inference footprint compared to those it’s competing with but equals Claude Opus 4.6 on the SWE-Bench Pro, which is a tough software engineering benchmark. It has been preferred by humans when evaluated side by side with Claude Sonnet 4.6 in blind tests. It gets 97% in AIME 2025, another benchmark test, and has a 256,000-token context window capable of loading a 600-page document into memory at once.
Also read: Computex 2026: Asus unveils Pro Art laptops powered by Nvidia RTX Spark chip
MAI-Code-1-Flash is the powerhouse behind everyday coding. It has been developed explicitly for the purpose of developer’s workflows and is already being integrated into GitHub Copilot for VS Code via the model and auto picker. Efficiency is its main advantage. Not only does it beat Claude Haiku 4.5 on every single coding benchmark that Microsoft tested it on but it uses up to 60% less tokens to do so. What’s more, it has been trained against the production harness of GitHub Copilot itself rather than just leaderboard benchmarks.
Neither is publicly available at the moment. The MAI-Thinking-1 model is in private preview on Microsoft Foundry, with a playground version set for release shortly. Meanwhile, MAI-Code-1-Flash is gradually becoming available for individual Copilot users.
What this means strategically is that, by deploying seven brand-new MAI models, Microsoft is creating a separate AI capability from OpenAI’s, independent of their roadmap, pricing, and agenda. It will remain to be seen whether these models would one day match the capabilities of anthro-pai and Google’s frontier-tier systems. But the groundwork is there already.
Also read: Computex 2026: Intel announces Xeon 6+ processors, says AI will make CPUs important again