Microsoft Launches 7 MAI Models at Build 2026: Direct Competition with Claude and Gemini

The AI Race Intensifies

While Microsoft had been heavily dependent on its partnership with OpenAI, there was always a question: when will the company develop its own models completely? At the Build 2026 conference, the answer came clearly and decisively. Microsoft AI, led by Mustafa Suleyman, announced the launch of 7 new AI models under the MAI brand.

This move is not just an addition to the list of available models. It is a strategic shift toward technical independence and direct competition with major companies in the field such as Anthropic and Google.

🔗 Official Source: microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models

The Hill-Climbing Machine Strategy

Before diving into the models, it is important to understand the philosophy behind them. Mustafa Suleyman, CEO of Microsoft AI, stated that the company is building what he calls a “Hill-Climbing Machine” – a system designed for continuous improvement and rapid iteration, relying on a massive increase in computational power.

The idea of hill-climbing means the company will not settle for releasing a model and then waiting. Instead, it will invest in infrastructure capable of periodically and steadily improving model performance, focusing on clean and legally licensed data, and moving away from the practice of distillation from other models.

The Seven Models: Complete Coverage of Tasks

Microsoft divided the MAI family into seven models, covering five main categories: reasoning and logic, coding, images, speech-to-text, and voice generation.

MAI-Thinking-1: The Leading Reasoning Model

This is the flagship model of the family and Microsoft’s first advanced reasoning model. It uses a Mixture-of-Experts architecture with 35 billion active parameters and a total of approximately one trillion parameters, with a context window of up to 256 thousand tokens.

Performance figures according to official numbers:

It achieved 53 percent on the SWE-bench Pro benchmark, a level comparable to Claude Opus 4.6. It achieved 97 percent on the AIME 2025 math problems benchmark and 94.5 percent on AIME 2026. In blind human side-by-side evaluations, judges preferred it over Claude Sonnet 4.6 in terms of overall quality.

The image below shows the Microsoft AI logo:

MAI-Code-1-Flash: The Ultra-Efficient Coding Model

This model is small in size compared to competitors, with only 5 billion active parameters, but its performance and reasoning power are impressive. It has been deeply integrated into Microsoft’s developer tools such as GitHub Copilot and VS Code. It achieved 51 percent on the SWE Bench Pro benchmark. In terms of cost, it is significantly cheaper than its competitors, with token consumption up to 60 percent lower on some tasks compared to other models.

MAI-Image-2.5: The Advanced Image Model

This model supports text-to-image generation and editing, and is available in two versions: the base version focused on high quality and the flash version focused on speed and efficiency for large-scale projects. In the Arena ELO ranking, it ranked third in image generation and second in image editing, surpassing Google’s Nano Banana Pro in this area. It has already been integrated into PowerPoint and OneDrive.

MAI-Transcribe-1.5: The Speech Transcription Model

Microsoft claims this is the most accurate model in the world for speech-to-text tasks, supporting 43 languages. Its speed is up to 5 times faster than competing models, and it supports specialized terminology in fields such as medicine and law. The price is highly competitive at only $0.36 per hour of transcription.

MAI-Voice-2: The Natural Voice Generation Model

This model generates natural and emotional speech from text and supports 15 languages. One of its most notable capabilities is voice adaptation, where it can mimic a specific person’s voice from a short audio sample, with protections against unauthorized use. It generates 60 seconds of audio in just one second.

The Real Innovation: Zero Distillation

What distinguishes MAI models from others is that they were trained from scratch on clean, licensed data, not randomly scraped from the internet. This approach is called Zero Distillation. Instead of using outputs from other models to train its own model, a common but legally and ethically controversial practice, Microsoft chose to build everything with its own data.

This means companies can use these models without fear of lawsuits related to copyright infringement.

Frontier Tuning: Unprecedented Model Customization

Alongside the models themselves, Microsoft announced Frontier Tuning. This is not just regular fine-tuning. It is a reinforcement learning process that occurs within your organization’s boundaries.

The idea is simple but revolutionary. Instead of using a general smart model that knows nothing about your company, you can train MAI-Thinking-1 on your specific business style, terminology, employee permissions, and even document formatting. Initial results were stunning. In one of Microsoft’s internal models, task completion rates jumped from 13 percent to 87 percent after applying Frontier Tuning. Microsoft also entered a partnership with Mayo Clinic to develop a leading medical model using this technology.

How to Access These Models?

All seven MAI models are available today through Azure AI Foundry. Additionally, they will be available on third-party platforms such as OpenRouter, Fireworks, and Baseten, giving developers flexibility in their choices. The models are also supported by Microsoft’s own Maia 200 processors, which offer 1.4 times better energy efficiency than other solutions.

Summary

Microsoft is no longer just an operator of OpenAI models. With the launch of MAI, the company has proven that it has the ability to compete in the most important race in Silicon Valley today. The focus on clean data and the hill-climbing strategy sets a new standard for transparency and efficiency, while Frontier Tuning redefines how companies benefit from generative AI.