FlagEval EmbodiedVerse - MMSI-Bench: leaderboard
BAAI FlagEval’s EmbodiedVerse suite, MMSI-Bench component: multi-image spatial intelligence, where the answer needs several views combined. Scored on FlagEval’s common scale so the thirteen component boards read against each other.
Metric: Score. Source: flageval.baai.ac.cn. Saturation forecast: Around December 2026. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 40.9 | #98 |
| 2 | Gemini 2.5 Pro | 39.2 | #152 |
| 3 | Qwen 3 VL 235B A22B Instruct | 33.8 | #273 |
| 4 | Qwen 2.5 VL 72B Instruct | 32.6 | #377 |
| 5 | Claude Sonnet 4.5 | 32 | #142 |
| 6 | Qwen 3 VL 235B A22B (Thinking) | 31.2 | #238 (Qwen 3 VL 235B A22B) |
| 7 | GPT-4.1 | 30.6 | #248 |
| 8 | Qwen 3 VL 30B A3B (Thinking) | 29.96 | #351 (Qwen 3 VL 30B A3B) |
| 9 | Qwen 3 VL 30B A3B Instruct | 29.4 | #384 |
| 10 | O4 Mini | 29.2 | #180 |
| 11 | Qwen 2.5 VL 32B Instruct | 28.3 | #453 |
| 12 | Qwen 2.5 VL 7B Instruct | 27.2 | #654 |
| 13 | Phi-4 Multimodal Instruct | 27.2 | #906 |
| 14 | InternVL3-8B | 27.1 | #620 |
| 15 | InternVL3-78B | 26.7 | #352 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=flageval-embodiedverse-mmsi-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.