MME
Multimodal · 2023-06-23
MME evaluates multimodal LLMs with 2,374 hand-written yes/no questions on 1,187 images (two per image) across 10 perception and 4 cognition subtasks. Each subtask scores accuracy plus accuracy+ (both questions on an image right) for up to 200 points, so the total runs from 0 to 2,800; answers are parsed by rule.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash-Lite | 2498 |
| InternVL2.5-78B | 2494 |
| Qwen3.5-4B | 2472 |
| Qwen2.5-VL-72B-Instruct | 2448 |
| LFM3-VL-2B | 2433 |
| Gemma 4 12B IT | 2400 |
| Qwen3.5-2B | 2349 |
| MiniCPM-V 2.6 | 2348 |
| Qwen2.5-VL-7B-Instruct | 2347 |
| GPT-4o (2024-05-13) | 2328 |
| Qwen2 VL 7B Instruct | 2327 |
| Gemma 4 E2B IT | 2229 |
| InternVL3-2B | 2186 |
| GPT-5.6 Luna | 2094 |
| LFM2-VL-3B | 2051 |
| LFM2.5-VL-3B | 2047 |
| GPT-4o Mini | 2003 |
| Claude 3.5 Sonnet (June 2024) | 1920 |
| Qwen2 VL 2B Instruct | 1872 |
| LFM2-VL-1.6B | 1757 |
| LFM2-VL-450M | 1230 |