PerceptionBench (GPT-6 Luna judge)
Multimodal
PerceptionBench's 3,000 atomic-perception questions with open answers graded by GPT-6 Luna instead of the benchmark's GPT-oss-120B grader, as Liquid AI ran them for its LFM3 comparison. Same items and construct as PerceptionBench, so it is a component of that family rather than independent evidence.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Luna | 48.8 |
| Gemma 4 12B IT | 43.5 |
| Qwen3.5-4B | 39.8 |
| Gemini 3.5 Flash-Lite | 37.7 |
| LFM3-VL-2B | 37.1 |
| Qwen3.5-2B | 29.5 |
| Gemma 4 E2B IT | 28.2 |