FMSU-Bench - Contextual Inference (Chinese): leaderboard
FMSU-Bench (2026) tests fine-grained speech understanding in Chinese and English with multiple-choice questions whose distractors include subtle acoustic confusions and text-only semantic traps, plus a tagged transcription task. This board is accuracy on identifying the context inferred from the vocal performance in the Chinese subset; percent, higher is better.
Metric: Accuracy (%) of identifying the context inferred from the vocal performance (CI), on the FMSU-Bench Chinese subset (500 to 1,000 manually verified items per task, from audiovisual sources), multiple-choice questions whose distractors include fine-grained acoustic confusions and text-only semantic traps; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 93.1 | #99 |
| 2 | Gemini 3 Flash | 86.9 | #175 |
| 3 | Gemini 2.5 Flash | 85.5 | #463 |
| 4 | Qwen 3 Omni 30B A3B | 80.1 | #847 |
| 5 | Qwen2.5-Omni-7B | 79.5 | #1884 |
| 6 | MiMo-Audio | 76.1 | #931 |
| 7 | FMSU-Bench Step-Audio 2 (checkpoint unspecified) | 69.4 | |
| 8 | Audio Flamingo 3 | 65.8 | |
| 9 | Kimi-Audio | 56 | #1735 |
| 10 | Qwen2-Audio (FMSU-Bench checkpoint unspecified) | 44.3 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=fmsu-bench-contextual-inference-chinese · How It Works · Data refreshed daily, snapshot 2026-11-21.