Microsoft Builds Its Own AI Models Instead of Just Using Others
Microsoft released two new artificial intelligence models today. This is part of a larger plan to build AI systems in-house rather than relying only on partnerships with other companies.
One model, called MAI-Thinking-1, is designed to tackle difficult problems that need careful reasoning—the kind of work that might normally require a human expert. The other, MAI-Code-1-Flash, helps programmers write code faster.
Both models were built using data that Microsoft owns the rights to, without borrowing from other companies' AI systems. For MAI-Thinking-1, the company left out data created by artificial intelligence itself. This approach matters because large companies increasingly want to know exactly where their AI's training data came from—especially for code-writing tools, where an AI might accidentally reproduce someone else's copyrighted code.
How Well Do These Models Work?
MAI-Thinking-1 was tested on challenging math problems and passed at a high rate (97% on one test, 94.5% on another). It also performs similarly to other leading reasoning models when tested on software engineering tasks. When human raters compared it side-by-side with competitors, they preferred MAI-Thinking-1, though Microsoft has not released detailed details about how this testing was done.
MAI-Code-1-Flash is built for speed. Programmers need code suggestions to arrive quickly when they are writing, and they often ask for multiple suggestions as they work. This model delivers faster results while using fewer computational resources than similar competing models.
Where You'll See These Models
MAI-Code-1-Flash is already being rolled out to individual users through GitHub Copilot—the AI tool that works inside the code editors many programmers use daily. This suggests Microsoft considers the model ready for real, everyday work rather than still in the testing phase.
Why Microsoft Is Building Its Own AI
For years, Microsoft relied heavily on partnerships with other AI companies, especially OpenAI. Now the company is building more AI systems itself. This shift started becoming clear in late 2024, when Microsoft announced plans to develop its own models for specific industries, alongside improvements to its AI tools and services.
The company also developed something it calls the Hill-Climbing Machine—a system for continuously improving models over time. Details are still limited, but it appears to be a framework for making models better through systematic adjustments.
The choice to use clean, commercially licensed data reflects what enterprise customers demand. Legal teams at large companies now routinely ask where an AI's training data came from, because they want to avoid using systems that were trained on stolen or copyrighted material. This is not a new technical problem—it is a business and legal one.
The decision to exclude AI-generated data addresses a different concern. Research has shown that if you train an AI model on data created by other AI models, the quality can degrade over time, like a photocopy of a photocopy. This pattern is called model collapse.
What This Means in the Broader Picture
Microsoft's push to develop its own AI models mirrors something we have seen before in technology. When cloud computing emerged, companies like Amazon and Microsoft started by renting access to software others had built. Over time, they built their own competing services. The outcome was usually better integration, lower costs, and more control—though it also meant taking on greater complexity and higher upfront spending.
In the AI landscape right now, good reasoning ability and code generation are features that enterprise customers expect. Microsoft's strategy focuses on practical, reliable integration rather than chasing the latest benchmark records. Enterprise buyers tend to value reliability and support over raw performance gains, and this release reflects that priority.
How These Models Fit Into Microsoft's Larger Plans
These new models can be used as building blocks in Microsoft Copilot Studio, where enterprises create and customize AI agents for their own needs. This suggests Microsoft sees these models not as standalone tools but as components in larger automated systems.
Microsoft is also working on AI for robotics through a model called Rho-alpha. While separate from today's announcements, it shows the company is developing AI that works across different types of tasks and media formats, not just text.
For developers and companies adopting AI, these models are solid incremental advances rather than major breakthroughs. Microsoft is optimizing for the ability to deploy and support AI at large scale, rather than pushing what is technically possible to its absolute limits. Whether this practical approach proves more valuable than competitors' focus on achieving the latest breakthroughs remains uncertain.
