Magenta: Closing the Loop Between Mathematical Reasoning and Lean Verification

Benchmark Model Rank Results
code-generation-on-scicodeK2-Horizon-375B-A23B–Accuracy: 42.7
code-generation-on-swe-bench-proK2-Horizon-375B-A23B–Resolved: 42.6
code-generation-on-terminal-bench-2-1K2-Horizon-375B-A23B–Accuracy: 70.2
mathematical-reasoning-on-aime-2025K2-Horizon-7B + Magenta–Accuracy: 100.0
multiple-choice-question-answering-mcqa-on-gpqa-diamondK2-Horizon-375B-A23B–Accuracy: 87.3
physics-question-answering-on-critptK2-Horizon-375B-A23B–Accuracy: 8.6