Audio Understanding on MMSU
55.7Perception ScoreUAS-Audio
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| UAS-AudioAudio Input Architecture=Continuous2026.04 | 55.7 | 77.4 | 66.2 | |
| MiMo-Audio-7B-Instruct + ThinkSize=7B, No LLM Modification=false, # Tokens=2.4T2026.03 | 51.7 | 74.8 | 62.9 | |
| EvASetting=Zero-shot2026.03 | 47.52 | 75.41 | — | |
| MiMo-Audio-7B-InstructSize=7B, No LLM Modification=false, # Tokens=2.4T2026.03 | 46.9 | 76.7 | 61.7 | |
| Gemini-1.5-Pro2026.03 | 46.3 | 76.1 | 60.7 | |
| Audio-Flamingo-3Setting=Zero-shot2026.03 | 45.63 | 77.86 | — | |
| Kimi-Audio-7B-InstructSetting=Zero-shot2026.03 | 45.47 | 71.85 | — | |
| ALARM-ESize=4B, No LLM Modification=true, # Samples=5.5M, # Hours=17K, # Tokens=1.5B2026.03 | 45.4 | 78.3 | 61.3 | |
| Kimi-AudioAudio Input Architecture=Continuous2026.04 | 44.8 | 75.7 | 59.8 | |
| Step-Audio-2-miniSetting=Zero-shot2026.03 | 44.36 | 78.32 | — | |
| Kimi-AudioSize=7B, No LLM Modification=false, # Hours=13.3M, # Tokens=585B2026.03 | 43.5 | 76 | 59.3 | |
| Step-Audio2Audio Input Architecture=Continuous2026.04 | 42.9 | 73.2 | 57.6 | |
| Step-Audio 2 miniSize=7B, No LLM Modification=false, # Tokens=0.2T2026.03 | 42.7 | 72.6 | 57.2 | |
| Qwen2.5-OmniAudio Input Architecture=Continuous2026.04 | 42.7 | 77.6 | 58.1 | |
| Qwen2.5-OmniSize=7B, No LLM Modification=false, # Tokens=0.3T2026.03 | 42.5 | 79.8 | 60.6 | |
| Qwen2.5-OmniSize=3B, No LLM Modification=false, # Tokens=0.3T2026.03 | 42.4 | 72.8 | 56.8 | |
| R1-AQASetting=Zero-shot2026.03 | 41.68 | 71.94 | — | |
| Audio-ReasonerSetting=Zero-shot2026.03 | 40.73 | 68.38 | — | |
| MiniCPM-o-2.6Size=7B, No LLM Modification=false2026.03 | 40.5 | 73.6 | 56.5 | |
| Qwen2.5-Omni-7BSetting=Zero-shot2026.03 | 40.37 | 66.31 | — | |
| GPT-4o Audio2026.03 | 39.7 | 71.2 | 56.4 | |
| ALARM-CASize=4B, No LLM Modification=true, # Samples=5.5M, # Hours=17K, # Tokens=1.5B2026.03 | 39.6 | 68.3 | 53.5 | |
| Qwen2-Audio-InstructSize=7B, No LLM Modification=false, # Hours=320K2026.03 | 39 | 68.9 | 53.3 | |
| ALARM-PSize=4B, No LLM Modification=true, # Samples=5.5M, # Hours=17K, # Tokens=1.5B2026.03 | 38.4 | 74.2 | 55.8 | |
| AL-Whisper-Instruct-RSize=4B, No LLM Modification=true, # Samples=0.5M, # Hours=1K, # Tokens=91M2026.03 | 38.2 | 73.6 | 55.3 | |
| UniAudio-TokenBackbone=Qwen2.5-3B, Tuning strategy=tuned identically2026.05 | 35.54 | 52.07 | 43.54 | |
| Baichuan-Omni-1.5Size=7B, No LLM Modification=false, # Hours=887K2026.03 | 35.4 | 67.2 | 50.6 | |
| Phi-4-MultimodalSize=4B, No LLM Modification=false, # Samples=> 28M, # Hours=> 2M2026.03 | 33.4 | 57.6 | 45 | |
| WavTokenizerBackbone=Qwen2.5-3B, Tuning strategy=tuned identically2026.05 | 32.83 | 45.37 | 38.9 | |
| Megrez-3B-OmniSize=2B, No LLM Modification=false2026.03 | 32.5 | 67.1 | 49 | |
| GLM-4-Voice-TokenizerBackbone=Qwen2.5-3B, Tuning strategy=tuned identically2026.05 | 32.4 | 47.64 | 39.78 | |
| StableTokenBackbone=Qwen2.5-3B, Tuning strategy=tuned identically2026.05 | 31.98 | 49.71 | 40.56 | |
| UAS-Audio-DAudio Input Architecture=Discrete2026.04 | 31.32 | 48.55 | 39.66 | |
| SALMONNSize=7B, No LLM Modification=false, # Samples=2.3M, # Hours=4.4K2026.03 | 29.8 | 30 | 30.1 | |
| CosyVoice2Backbone=Qwen2.5-3B, Tuning strategy=tuned identically2026.05 | 27.44 | 45.83 | 36.34 | |
| LTUSize=7B, No LLM Modification=false, # Samples=5.6M2026.03 | 20.8 | 24.4 | 22.6 | |
| GLM-4-VoiceAudio Input Architecture=Discrete2026.04 | 11.04 | 16.16 | 13.3 | |
| Gemini-3.1 Pro2026.04 | — | — | 81.3 | |
| Qwen3.5-Omni-Flash2026.04 | — | — | 72.2 | |
| Qwen3.5-Omni-Plus2026.04 | — | — | 82.8 |