Unified Audio Intelligence Without Regressing on Text Intelligence

Benchmark Model Rank Results
agents-on-benchAudex 30B-A3B–Accuracy: 57.2
audio-generation-on-audiocapsAudex-30B-A3B SFT–FD_openl3: 60.9
audio-understanding-reasoning-on-mmarAudex 30B-A3B–Accuracy: 63.2
audio-understanding-reasoning-on-mmsuAudex 30B-A3B–Accuracy: 63.4
code-generation-on-livecodebench-v6Audex 30B-A3B–Pass@1: 85.3
code-generation-on-scicodeAudex 30B-A3B–Accuracy: 35.2
code-generation-on-swe-bench-verifiedAudex 30B-A3B (OpenHands)–Accuracy: 48.2
code-generation-on-terminal-bench-2-0-terminus-2Audex 30B-A3B (Terminus-2)–Accuracy: 19.1
instruction-following-on-arena-hard-auto-v2Audex 30B-A3B–Accuracy: 81.6
instruction-following-on-ifbenchAudex 30B-A3B–Accuracy: 77.8
mathematical-reasoning-on-aime-2025Audex 30B-A3B–Accuracy: 91.2
mathematical-reasoning-on-aime-2026Audex 30B-A3B–Accuracy: 89.4
mathematical-reasoning-on-hmmt-feb-2025Audex 30B-A3B–Accuracy: 92.2
multi-task-language-understanding-on-mmlu-proAudex 30B-A3B–Accuracy: 78.9
multi-task-language-understanding-on-mmlu-reduxAudex 30B-A3B–Accuracy: 86.4
multiple-choice-question-answering-mcqa-on-gpqa-diamondAudex 30B-A3B–Accuracy: 74.9
speech-recognition-on-librispeech-test-cleanAudex 30B-A3B–Word Error Rate (WER): 1.34
speech-recognition-on-librispeech-test-otherAudex 30B-A3B–Word Error Rate (WER): 3.06