LLM Stats (MME): leaderboard
MME perception-plus-cognition total, expressed on a common 0–100 scale from the benchmark’s 0–2800 sum. The five reviewed LLM Stats observations retain their cited vendor evaluation protocols; DeepSeek-VL2 report totals are restored from the relay’s mis-scaled values and divided by 28, while Liquid’s already-normalized published values remain unchanged. LFM3-VL uses its explicitly reported thinking configuration. This is normalized MME score, not ordinary accuracy percent.
Metric: MME total normalized to 0–100. Source: llm-stats.com. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LFM3-VL-2B (Thinking) | 86.9 | |
| 2 | DeepSeek VL2 | 80.46 | |
| 3 | DeepSeek VL2 Small | 75.82 | |
| 4 | LFM2.5-VL-3B | 73.1 | |
| 5 | DeepSeek VL2 Tiny | 68.39 |
Interactive version: theaggregate.ai/benchmark?slug=llm-stats-mme · How It Works · Data refreshed daily, snapshot 2026-11-05.