Emu3-Chat: benchmark results
BAAI Emu3 tokenizes multimodal inputs into a discrete vocabulary and trains an autoregressive transformer from scratch using next-token prediction, with a separately loaded vision tokenizer. The Chat checkpoint supports image-conditioned text responses and remains separate from the Gen and first-stage pretrained releases. Provider: BAAI. Released 2024-09-27. Access: Open.
Aggregate Rating 1285 ± 28, rank #8913 of 10916 rated models, from 519 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| OpenVLM MMT-Bench - Spot The Diff | 100 | Score (%) | 99.5 |
| OpenVLM MMT-Bench - MIA | 82.5 | Score (%) | 97.1 |
| OpenVLM MMT-Bench - Multiple View Image Understanding | 40 | Score (%) | 89.3 |
| OpenVLM MMT-Bench - Rotated Object Detection | 66.7 | Score (%) | 88.6 |
| OpenVLM MMT-Bench - Salient Object Detection RGBD | 55 | Score (%) | 85.2 |
| OpenVLM MME - Position | 158.3 | Score | 84.2 |
| OpenVLM MMT-Bench - 3D Indoor Recognition | 40 | Score (%) | 82.5 |
| OpenVLM MMT-Bench - Animal Keypoint Detection | 40 | Score (%) | 82 |
| OpenVLM MMT-Bench - Jigsaw Puzzle Solving | 35 | Score (%) | 81.6 |
| OpenVLM MMT-Bench - Counting By Reasoning | 85 | Score (%) | 81.3 |
| OpenVLM MMT-Bench - Scene Text Recognition | 95 | Score (%) | 81.3 |
| OpenVLM MMT-Bench - Scene Emotion Recognition | 60 | Score (%) | 80.6 |
Interactive version: theaggregate.ai/model?slug=emu3-chat · How It Works · Data refreshed daily, snapshot 2026-11-08.