M3Exam (Multimodal Memory): leaderboard
M3Exam (2026) is a multimodal conversational-memory benchmark of 5,150 questions over long multi-session user-assistant conversations with images and documents; the frontier models here answer with the gold supporting sessions in context. The headline is an overall score (0-100) that puts half its weight on an LLM-judge score and splits the rest over graded exact match, token F1 and BLEU-1; higher is better.
Metric: Overall Score (0-100): item-weighted composite with half weight on the LLM-judge score and the other half split over graded exact match, token F1 and BLEU-1, on 5,150 questions over 239 long multi-session user-assistant conversations with images and PDFs (15 personas); each model answers with the annotated gold supporting sessions and their images or rendered pages placed in context, no memory module; printed 0-1 and x100 here; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.1 | 54.93 | #179 |
| 2 | Qwen 3.6 Plus | 50.49 | #222 |
| 3 | Seed 2.0 Pro | 49.38 | #211 |
| 4 | Claude Opus 4.6 | 49.37 | #102 |
| 5 | GPT-5.4 | 48.23 | #123 |
| 6 | Gemini 3.1 Pro (Preview) | 46.68 | #94 |
| 7 | Kimi K2.5 | 44.93 | #256 |
| 8 | Grok 4 | 43.57 | #307 |
| 9 | Qwen 2.5 VL 7B Instruct | 41.79 | #1833 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=m3exam-multimodal-memory · How It Works · Data refreshed daily, snapshot 2026-11-13.