Microsoft Strikes Out On Its Own

2026-07-03 · Source: The Batch | DeepLearning.AI | AI News & Insights - www.deeplearning.ai · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Advanced, short

Summary

Microsoft unveiled MAI-Thinking-1, its inaugural reasoning language model developed entirely from scratch, marking a departure from its previous reliance on fine-tuning or distilling third-party models. This medium-sized model, comparable to Claude Sonnet 4.6, leads a family of seven MAI models, including MAI-Code-1-Flash for coding, introduced at Microsoft's Build conference. MAI-Thinking-1 features a Mixture of Experts architecture with 1 trillion total parameters (35 billion active per token) and supports up to 256,000 tokens for both input and output. Its capabilities include function calling and developer instructions, with compatibility for OpenAI's Chat Completions API. Microsoft's tests show it achieved 97.0 percent on the AIME 2025 math benchmark, surpassing Claude Sonnet 4.6 (95.6 percent) and DeepSeek V3.2 (93.1 percent), though it trailed Claude Opus 4.6 (99.8 percent). The model was trained on 30 trillion pretraining tokens and 3.55 trillion midtraining tokens, primarily human-generated data, including over 50 percent code, and is currently in private preview.

Key takeaway

For AI Engineers evaluating new reasoning models, MAI-Thinking-1 offers a compelling option within the Microsoft ecosystem. You can now access a capable, directly trained model without adding new vendors or moving data outside your existing Azure and Copilot tools. This simplifies integration and potentially reduces vendor complexity, especially if your team prioritizes models with steerable behavior from direct training. Explore its private preview to assess its fit for your specific mathematical or reasoning tasks.

Key insights

Microsoft developed MAI-Thinking-1 from scratch for steerable behavior, avoiding third-party model inheritance.

Principles

• Training directly yields more steerable model behavior.
• Avoid synthetic data to prevent inheriting teacher's design.
• Reinforcement learning consolidates specialist model capabilities.

Method

Pretrain a base model, fine-tune specialists, distill into a student, then apply reinforcement learning for reasoning and safety.

In practice

• Utilize developer instructions to prioritize system prompts.
• Integrate via OpenAI's Chat Completions API for compatibility.
• Consider MAI-Code-1-Flash for coding assistance.

Topics

• MAI-Thinking-1
• Reasoning Models
• Mixture-of-Experts
• Reinforcement Learning
• Model Training
• Microsoft Azure
• Function Calling

Best for: Investor, CTO, VP of Engineering/Data, AI Scientist, AI Engineer, Director of AI/ML

Related on AIssential

• Microsoft Says MAI-Thinking-1 Was Not Distilled From Another LLM — Microsoft's MAI-Thinking-1 demonstrates competitive LLM performance is achievable without relying on synthetic data or third-party model distillation.
• Why Microsoft Trained MAI-Thinking-1 Without Synthetic Data — Microsoft's MAI-Thinking-1 achieves competitive reasoning performance by strictly avoiding synthetic data and actively removing AI-generated content during pre-training.
• Microsoft Build 2026: Be yourself at work — Microsoft's strategy for agentic AI centers on providing a diverse, open, and secure platform for developers.
• 🤖 AI Agents Weekly: Microsoft's Seven MAI Models, Gemma 4 12B, NVIDIA Nemotron 3 Ultra, Agents' Last Exam, Devin Desktop, and More — New multimodal AI models are enabling advanced agentic reasoning and local deployment on consumer hardware.
• [AINews] Microsoft Build: MAI-Thinking-1 and MAI Family models — Microsoft's MAI models showcase a transparent, full-stack AI strategy emphasizing clean data and local agent execution.
• LAI #131: A Tool Call Can Succeed and Still Be the Wrong Tool — Microsoft's MAI models prioritize human-generated data and transparent lineage over synthetic data for foundational capabilities.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Batch | DeepLearning.AI | AI News & Insights - www.deeplearning.ai.