Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters

Benchmark Model Rank Results
mathematical-reasoning-on-hmmt-feb-2025Step 3.5 Flash (11B)#1Accuracy: 98.4
agents-on-benchStep-3.5-Flash#2Accuracy: 88.2
mathematical-reasoning-on-aime-2025Step-3.5-Flash#2Accuracy: 97.3
instruction-following-on-multichallengeStep-3.5-Flash PaCoRe#5Accuracy: 60.8
instruction-following-on-multichallengeStep-3.5-Flash#9Accuracy: 55.7
multi-task-language-understanding-on-mmluStep-3.5-Flash Base#7Average (%): 85.8
long-context-understanding-on-longbench-v2Step-3.5-Flash#10Accuracy: 57.5
question-answering-on-humanity-s-last-exam-hleStep-3.5-Flash PaCoRe#16Accuracy: 27.9
question-answering-on-humanity-s-last-exam-hleStep-3.5-Flash#22Accuracy: 23.1
sentence-completion-on-hellaswagStep-3.5-Flash Base#16Accuracy: 90.2
multi-task-language-understanding-on-mmlu-proStep-3.5-Flash PaCoRe#18Accuracy: 84.8
multi-task-language-understanding-on-mmlu-proStep-3.5-Flash#20Accuracy: 84.4
multi-task-language-understanding-on-mmlu-proStep-3.5-Flash Base#90Accuracy: 62.3
instruction-following-on-ifbenchStep-3.5-Flash#22Accuracy: 67.4
instruction-following-on-ifbenchStep-3.5-Flash PaCoRe#32Accuracy: 56.8
code-generation-on-swe-bench-multilingualStep-3.5-Flash#25Accuracy: 67.4
code-generation-on-swe-bench-verifiedStep-3.5-Flash#26Accuracy: 74.4
deep-research-agents-on-browsecompStep-3.5-Flash (w/ context manager)#29Accuracy: 69.0
deep-research-agents-on-browsecompStep-3.5-Flash#41Accuracy: 51.6
code-generation-on-mbppStep-3.5-Flash Base#36Accuracy: 79.4
common-sense-reasoning-on-winograndeStep-3.5-Flash Base#36Accuracy: 79.1
multiple-choice-question-answering-mcqa-on-gpqa-diamondStep-3.5-Flash PaCoRe#38Accuracy: 85.0
multiple-choice-question-answering-mcqa-on-gpqa-diamondStep-3.5-Flash#44Accuracy: 83.5
math-word-problem-solving-on-mathStep-3.5-Flash Base#40Accuracy: 66.8
code-generation-on-humanevalStep-3.5-Flash Base#56Pass@1: 81.1
arithmetic-reasoning-on-gsm8kStep-3.5-Flash Base#90Accuracy: 88.2