MAI-Thinking-1 operates like a methodical problem-solver that breaks complex challenges into careful steps before committing to an answer. Its MoE architecture means it activates only the parameters it needs, keeping responses fast and costs low despite its massive total parameter count. It excels at multi-step reasoning, long-context analysis, and code generation, matching heavier models on difficult benchmarks while maintaining practical efficiency. Particularly strong on mathematical reasoning and software engineering tasks.
Frontier-level reasoning: 97% AIME 2025, 94.5% AIME 2026.
53% SWE-Bench Pro, matching Claude Opus 4.6.
Function calling support. Preferred over Sonnet 4.6 in blind evals.
Primarily English-focused reasoning model.
| Benchmark | Score | Type | Recorded |
|---|---|---|---|
| AIME 2025 | 97.0 | percentage | 1mo ago |
| SWE-Bench | 53.0 | percentage | 1mo ago |