Nanbeige4.2-3B: Unlocking Agentic Capabilities in a Compact Model

Benchmark Model Rank Results
code-generation-on-terminal-bench-2-0-terminus-2Nanbeige4.2-3B (Terminus-2)#7Accuracy: 44.1
mathematical-reasoning-on-hmmt-feb-2026Nanbeige4.2-3B#7Accuracy: 82.8
code-generation-on-scicodeNanbeige4.2-3B#18Accuracy: 35.6
multiple-choice-question-answering-mcqa-on-gpqa-diamondNanbeige4.2-3B#22Accuracy: 87.4
code-generation-on-livecodebench-v6Nanbeige4.2-3B#29Pass@1: 72.5
question-answering-on-humanity-s-last-exam-hleNanbeige4.2-3B (w/o search)#33Accuracy: 17.8
instruction-following-on-ifbenchNanbeige4.2-3B#34Accuracy: 54.6
code-generation-on-swe-bench-proNanbeige4.2-3B#37Resolved: 46.9
code-generation-on-swe-bench-verifiedNanbeige4.2-3B (OpenHands)#57Accuracy: 63.6