project
MAI-Thinking-1 - Microsoft's first self-developed advanced inference model
MAI-Thinking-1 is Microsoft's first self-developed advanced inference model. The model employs a sparse MoE architecture with 35B active parameters and approximately 1T total parameters, and was trained entirely from scratch on clean, commercially licensed data, without using any...
What is MAI-Thinking-1?
MAI-Thinking-1 is Microsoft's first self-developed advanced inference model. The model employs a sparse MoE architecture with 35B active parameters and approximately 1T total parameters, and was trained entirely from scratch on clean, commercially licensed data without using any third-party model distillation. It achieves industry-leading performance on key benchmarks such as SWE-Bench Pro and AIME 2025/2026, matching Claude Opus 4.6 and outperforming Sonnet 4.6 in blind tests.
Main functions of MAI-Thinking-1
-
Advanced Mathematics and Scientific ReasoningIt achieved 97.0% on AIME 2025 and 94.5% on AIME 2026, demonstrating extremely strong STEM reasoning capabilities.
-
Intelligent Software EngineeringIt achieved a score of 52.8% on SWE-Bench Pro and 87.7% on LiveCodeBench v6, supporting multi-step code reading, editing, testing, and error recovery.
-
Long context processingSupports context windows up to 256K, capable of handling long documents and large-scale codebases.
-
Tool Invocation and Agent CapabilitiesIt can call external tools to interact with the environment and perform complex, multi-step tasks.
-
Balance between safety and usefulness: Optimize helpfulness and safety simultaneously during RL training, and reduce harmful output.
Technical Principles of MAI-Thinking-1
- Sparse MoE architectureMAI-Thinking-1 employs a 78-layer Decoder-only Transformer, alternating between sparse MoE layers with 8/512 expert activations and dense FFN layers, along with a 5:1 local/global attention mechanism. This achieves performance close to 1T total parameters while maintaining 35B active parameters, significantly reducing KV cache and computational overhead during inference.
- Data and pre-trainingThe model is pre-trained from scratch on clean, commercially licensed data with 30T tokens, strictly excluding AI-generated content and open-source training sets, and decontaminating common benchmarks; mid-term training further enhances STEM, mathematical, and coding capabilities, and ultimately supports ultra-long contexts of up to 256K.
- Strengthening learning to "climb the hill"Microsoft proposed the concept of "Hill-Climbing Machine," which learns the inference chain from scratch using a self-developed RL framework. It does not rely on prior CoT distillation and performs thousands of steps of continuous log-linear improvement in three areas: STEM inference, Agentic coding, and Helpfulness & Safety.
- Evaluation systemInternally, nearly 40 NLL benchmarks have been established covering five major categories: code, STEM, mathematics, general knowledge, and multilingualism. Negative log-likelihood non-multiple-choice questions are given priority for evaluation to avoid formatting bias. At the same time, private benchmarks have been built to prevent data contamination and ensure that the evaluation results truly reflect the model's capabilities.
How to use MAI-Thinking-1
The specific usage methods and access channels for MAI-Thinking-1 have not yet been announced.
MAI-Thinking-1's core advantages
-
Completely self-developed without distillationIt does not inherit the design biases of third-party models, and has stronger controllability and adaptability.
-
Medium size, top performanceWith only 35 active parameters, it can be compared to larger models such as Claude Opus 4.6, resulting in lower inference costs and more flexible deployment.
-
Data is clean and traceableExclude AI-generated content to ensure that model behavior is explainable and improveable.
-
Full-stack self-controlFrom Microsoft's self-developed accelerator to the training framework and RL environment, everything is built internally and can be optimized end-to-end.
-
Continuous evolution capabilityThe ramp-up machine architecture supports the absorption of better data, stronger rewards, and more computing power, enabling continuous improvement in capabilities.
MAI-Thinking-1 project address
- Project official websitehttps://microsoft.ai/news/introducing-mai-thinking-1/
- Technical Papers: https://microsoft.ai/wp-content/uploads/2026/06/main_20260602_2.pdf
Comparison of MAI-Thinking-1 with similar competing products
| Dimension | MAI-Thinking-1 | Claude Opus 4.6 | DeepSeek-R1 |
|---|---|---|---|
| Architecture | 35B Active / 1T Total Parameters MoE | Not publicly available (large-density model) | 37B Active / 671B Total Parameters MoE |
| Training methods | Training from scratch, without distillation | Not disclosed | Distillation based on DeepSeek-V3 |
| SWE-Bench Pro | 52.8% | ~52.8% (unchanged) | Approximately 49-50% |
| AIME 2025 | 97.0% | Unclear | Approximately 79.8% |
| Context length | 256K | 200K | 128K |
| Data Strategy | Pure commercial data, no AI-generated content. | Unclear | Open source + synthetic data |
| Full-stack autonomy | Yes (Microsoft's self-developed hardware and framework) | No (Anthropic) | Partially dependent on the NVIDIA ecosystem. |
Application scenarios of MAI-Thinking-1
-
Enterprise software developmentIt assists developers in code review, bug fixing, feature implementation, and automated testing.
-
Scientific Research and Complex ComputingSolve complex reasoning problems in advanced mathematics, physics, chemistry, and other fields.
-
Long document analysis and knowledge managementDeep understanding and summarization of extremely long texts such as legal contracts, technical documents, and academic papers.
-
AI Agent WorkflowAs the core inference engine, it drives Agent tasks such as automated office work, data analysis, and system maintenance.