MCGA: leaderboard
MCGA (Du et al., Harbin Institute of Technology, 2026) tests audio LLMs on recordings of classical Chinese literature (shi, ci, qu, fu and wen) read by native speakers, across six tasks with 1,000 test recordings each: recognition, translation into English, emotion captioning, open-ended spoken QA, and multiple-choice understanding and reasoning. The Score adds the six task scores (recognition counted as 100 minus the character error rate) on a 0-600 scale; higher is better.
Metric: Score (0-600; sum of the six task scores on 1,000 test recordings each, recognition counted as 100 minus CER). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3 Omni 30B A3B Instruct | 442.9 | #902 |
| 2 | Gemini 3 Flash | 440.9 | #159 |
| 3 | Qwen2.5-Omni-7B | 380.7 | #1797 |
| 4 | Voxtral-Small-24B-2507 | 292.9 | #1291 |
| 5 | Voxtral-Mini-3B-2507 | 244.7 | #3702 |
| 6 | Phi-4 Multimodal Instruct | 210.1 | #3914 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mcga · How It Works · Data refreshed daily, snapshot 2026-11-12.