Microsoft Builds Its Own AI Models — Here's What It Means

Microsoft released two new AI models today as part of a bigger shift: the company is now building advanced AI systems in-house instead of relying only on partnerships with external developers.

The two new models are called MAI-Thinking-1 and MAI-Code-1-Flash. The first is designed to solve complex problems step-by-step, much like asking an expert to work through a difficult math problem. The second is built for rapid code generation — helping software developers write code faster and with fewer errors.

What sets these models apart is how they were trained. Microsoft built them without relying on AI-generated training data from other companies, and used only data the company has commercial rights to use. This matters because large corporations are increasingly concerned about where their AI models' training data comes from, especially for code generation tools that might accidentally copy someone else's work.

How These Models Perform

MAI-Thinking-1 uses a technique called Mixture of Experts, which is like hiring a team of specialists where only the relevant experts work on each problem, rather than putting every specialist to work on every task. This approach lets the model be huge (about 1 trillion total parameters) but only use 35 billion parameters at a time, which makes it faster and more efficient.

On standardized tests for math reasoning, the model scores 97.0% and 94.5% on different difficulty levels — comparable to the leading reasoning models available today. For software engineering benchmarks, Microsoft reports it matches top competitors.

MAI-Code-1-Flash takes a different approach, optimized purely for speed. It outperforms competing models on code generation while using fewer tokens — think of tokens as individual words or code chunks. For developers who test multiple code suggestions in rapid succession, fewer tokens means faster results and lower costs.

Where You'll See These Models

MAI-Code-1-Flash is already showing up in GitHub Copilot, the AI coding assistant that works inside Visual Studio Code. It's rolling out to individual users now and appears as an option when you want AI code suggestions. This suggests Microsoft views it as production-ready for everyday developer work, not an experimental tool.

Why Microsoft Is Building Its Own AI

The broader shift here is strategic. In November 2024, Microsoft announced a suite of models tailored to specific industries, alongside improvements to its Azure AI infrastructure. Rather than always licensing models from OpenAI or other partners, Microsoft is investing heavily in developing its own.

This pattern has played out in the tech industry before. When cloud computing emerged, major companies like Amazon and Microsoft initially resold software from other vendors. Over time, they built their own competing services. The result: better integration between services, lower costs, and more control over development roadmaps. But it also required more engineering investment and added complexity.

Microsoft's emphasis on "clean, commercially licensed data" reflects what enterprise customers are demanding. Legal teams at large corporations want to know exactly where their AI models' training data came from. This is partly about avoiding legal risk — a code generation tool that accidentally reproduces copyrighted code could be a problem — and partly about understanding what you're buying.

The decision to exclude AI-generated content from MAI-Thinking-1's training addresses a technical concern called model collapse. When you train an AI model on data created by other AI models, quality can gradually degrade across generations, like making a photocopy of a photocopy. Researchers have shown this is a real risk as AI-generated data becomes more common in training datasets.

What This Means for Users and Developers

For developers and enterprise AI teams, these new models represent solid improvements in areas where Microsoft sees demand — complex reasoning and rapid code generation. They're not revolutionary, but they're practical.

Microsoft's strategy here seems to emphasize reliability and integration over chasing the latest benchmarks. The company is optimizing for broad adoption at scale rather than winning headline performance races. For enterprises choosing AI tools, this approach often matters more than raw capability numbers, since support, integration with existing systems, and stability in production use tend to drive purchasing decisions. Whether this pragmatic approach proves more valuable than competitors' focus on breakthrough capabilities is something we'll learn over the coming months.