Instead of just throwing massive compute at the wall and hoping for the best, the authors of the paper approached the entire process as a system-level optimization problem with a setup they call a hill-climbing machine. The end result is a Mixture-of-Experts model with 35 billion active parameters and 1 trillion total parameters. It performs incredibly well on tough math and coding benchmarks, scoring 97.0% on AIME 2025 and 52.8% on SWE-Bench Pro.
They trained MAI-Thinking-1 entirely from scratch using 30 trillion tokens of clean human-generated data and explicitly avoided distilling data from other frontier models or using scrubbed AI-generated content from their datasets. The team argues that while copying other models is faster, building capabilities from the ground up creates a more robust and steerable system. To handle this massive undertaking, they built a custom training framework called YOLO to run on up to 8000 GB200 GPUs which allowed them to control everything from the hardware kernels to the network communications.
Instead of trying to teach one model everything at once, they branched the training into three separate paths. One specialist for math and science, one for agentic coding and tool use, and one strictly for safety and helpfulness. Once these specialists were fully baked, they merged their knowledge together using a supervised fine-tuning process. A final lightweight reinforcement learning run polished the consolidated agents in to the final model.
The model holds its own against Claude Sonnet 4.6 across various benchmarks, and they also ran extensive human evaluations to ensure the model was actually helpful in real-world scenarios rather than just optimized for automated tests.