Microsoft unveils MAI-Base-2, an AI model with 2.8 trillion parameters that outperforms competitors

Microsoft has completed the pre-training phase of its latest artificial intelligence model, MAI-Base-2, which boasts 2.8 trillion parameters. The training process utilized 16,000 Nvidia GB200 GPUs over 38 days, spanning August and September. According to Microsoft, internal tests showed that the model outperformed open-source alternatives in programming, Q&A, science, and math tasks.

This launch reinforces Microsoft’s strategy of developing in-house solutions for tasks that could also be handled by partners like OpenAI and Anthropic. The MAI family is already integrated into products such as Excel, Outlook, and GitHub Copilot, with the aim of increasing control over the technology used.

Why Microsoft builds its own models

During Microsoft’s July earnings call, CEO Satya Nadella compared MAI-Code-1-Flash to GPT-5.6 in common Excel tasks, noting that Microsoft’s model delivers comparable quality at a significantly lower cost. Nadella also stressed that diversifying models helps ensure service continuity if a supplier or technology needs to be changed. “Every model is replaceable,” he said during an investor conference call.

This approach is particularly important in the corporate market, where Copilot is focused. In this sector, model selection depends on execution cost, integration with existing systems, and the ability to deliver results that meet business needs.

Architecture and parameters of MAI-Base-2

MAI-Base-2 uses a mixture of experts (MoE) architecture, with 80 billion parameters active. In this model, specific components are selected to process each input, allowing total capacity to be expanded without activating all parameters at every stage. Other labs, such as Mistral Large 4, also use this architecture.

While the number of parameters is impressive, performance does not depend solely on this factor. Architecture, data quality, training, and the nature of the task being evaluated also influence results.

Partnership with OpenAI remains active

Internal development does not mean ending collaboration with OpenAI, the creator of ChatGPT. An agreement updated in April renewed the intellectual property license for OpenAI’s models and products until 2032, now on a non-exclusive basis. Microsoft remains OpenAI’s primary cloud partner and shareholder, allowing it to maintain cooperation while exploring other business opportunities.

After pre-training, MAI-Base-2 is undergoing a post-training phase, where fine-tuning adjustments are being made. Microsoft plans to license the model’s weights to select partners through Foundry, its platform dedicated to AI development and use.

Frequently Asked Questions

• What is MAI-Base-2? It is a foundational artificial intelligence model developed by Microsoft, whose pre-training was completed and announced in October 2026.
• Is it available to the public yet? The model is still in the post-training phase and has not been released for external use.
• Do more parameters mean a better AI? Not necessarily. Performance also depends on data, architecture, training, and the task being evaluated.
• Did the model outperform ChatGPT? The announcement does not claim this. The comparisons released refer to open-source models tested internally.
• Has Microsoft ended its partnership with OpenAI? No. The agreement updated in April maintains the collaboration and the intellectual property license until 2032, now on a non-exclusive basis.