Details
- Microsoft AI announced MAI-Thinking-1, its first end-to-end in-house reasoning model designed for STEM reasoning and coding tasks.
- The model uses a 35B active / 1T total parameter mixture-of-experts (MoE) architecture, indicating substantial sparsity and scale.
- On coding, MAI-Thinking-1 scores 52.8% on SWE-Bench Pro, which Microsoft describes as competitive with Anthropic Claude Opus 4.6 on this benchmark.
- On math reasoning, the model attains 97% on the AIME 2025 benchmark, placing it among high-performing frontier reasoning models.
- Microsoft cites a pre-training strategy centered on simple scaling laws and data-driven iteration over architecture and data, emphasizing a fully clean data lineage with in-house processed data.
- The reinforcement learning (RL) phase starts from a checkpoint with no prior exposure to reasoning traces, avoiding distillation from larger models' chains-of-thought and focusing on learned, not inherited, capabilities.
- The RL "climb" trains the model to use chains of thought, tools, and interactive environments, while aligning with preference and safety signals.
- Microsoft positions MAI-Thinking-1 as the first in a series of reasoning-focused models, underpinned by internal recipes for scaling RL-based reasoning from a clean, human-written data foundation.
Impact
MAI-Thinking-1 signals Microsoft’s push to own the full stack of reasoning model development rather than relying on distillation from external frontier models. By combining a large MoE backbone with RL that begins from a reasoning-trace-free checkpoint, Microsoft is testing whether cleaner data lineage and learned reasoning can close the gap with leading models in coding and STEM, potentially intensifying competition with OpenAI, Anthropic, and Google on specialized reasoning benchmarks.
