Microsoft announced MAI-Thinking-1, the company’s first reasoning language model that it built independently rather than by distillation or fine-tuning from another developer’s model. The model was presented at Microsoft’s Build conference and is described by Microsoft as a medium-sized model comparable to Claude Sonnet 4.6.
Technical specs and capabilities
- Input/output: text in and out, up to 256,000 tokens.
- Architecture: mixture of experts (MoE) — 1 trillion total parameters, with 35 billion parameters active per token.
- Features: function calling, developer instructions (these are prioritized above user prompts in conflicts), and compatibility with OpenAI’s Chat Completions API.
How it was trained
Microsoft described a multi-stage training pipeline: pretraining a base model, fine-tuning separate copies into specialist “teacher” models, distilling those into a student model, and then using reinforcement learning to teach the student to reason.
The company reported approximately 30 trillion tokens for pretraining and 3.55 trillion tokens for midtraining, mostly human-generated data and with over 50% of the training data consisting of code. Post-training data included more than 5 million STEM questions and over 160,000 coding questions.
Microsoft said it relied primarily on licensed material for pretraining and avoided using synthetic (AI-generated) data as a principal source. The stated rationale is that models trained on AI-generated data or distilled from third-party teachers risk inheriting those teachers’ design choices and that direct training yields more steerable behavior.
For web data, Microsoft crawled roughly 1.2 trillion webpages, filtered that set down to 794 billion pages, and additionally incorporated 24.2 billion deduplicated webpages from Common Crawl.
Specialist models and consolidation
Using reinforcement learning on the base model, the team trained three specialists: one for STEM reasoning, one for agentic coding and tool use, and one focused on helpfulness and safety. Rather than imitating reasoning traces from other models, these specialists generated original chains of thought and were rewarded for correct math solutions, passing code test cases, and sound judgment.
The three specialists were consolidated in two stages: (i) supervised fine-tuning to distill them into a single model and (ii) a further reinforcement learning round to reduce over-refusals and improve safety and style.
Performance
According to Microsoft’s internal tests, MAI-Thinking-1 is strongest on mathematics and performs less well on graduate-level science and agentic coding relative to some competitors. On the AIME 2025 benchmark (competition math problems) Microsoft reported MAI-Thinking-1 at 97.0%, ahead of Claude Sonnet 4.6 (95.6%) and DeepSeek V3.2 (93.1%), but behind Claude Opus 4.6 (99.8%). No independent evaluations or comparisons to newer models have been published so far.
Availability and undisclosed details
MAI-Thinking-1 is available in a private preview through Microsoft Foundry; Microsoft plans broader access later via Fireworks AI, Baseten, and OpenRouter. The company has not disclosed the full list of datasets or data providers, nor pricing.
The MAI family announced at Build includes seven models, among them MAI-Code-1-Flash, a smaller coding model that will be available in GitHub Copilot and Visual Studio Code.
Background and significance
Historically, Microsoft relied heavily on OpenAI’s models for products such as Copilot and previously built models drawing on rivals’ work: the Phi family distilled OpenAI’s GPT-4 and GPT-5, and MAI-DS-R1 was a fine-tuned version of DeepSeek-R1. In April 2026 Microsoft and OpenAI amended their partnership so Microsoft’s license to OpenAI models became non-exclusive, allowing OpenAI to serve products on other cloud providers.
An independently trained reasoning model on Microsoft’s stack means teams already using Microsoft tools may access a capable reasoning model without onboarding an external vendor or moving data out of their existing environment, which could attract Azure and Copilot customers. Microsoft says it plans additional models built on the MAI training pipeline.
Open questions
Microsoft did not publish the complete dataset list, exact dataset compositions, or pricing. While the company states it prioritized licensed and human-generated data, it also incorporated large amounts of web data — a common practice across major model developers.
Overall, MAI-Thinking-1 signals Microsoft’s move toward self-developed reasoning capabilities and lays the groundwork for further models based on the same data pipeline.
