LFM3-VL-2B Release - ActivityNet-QA (test): leaderboard
Liquid AI LFM 3 vision release card; vLLM 0.27.1 and recommended generation settings; reasoning enabled except LFM 2.5-VL (unsupported) and explicit no-thinking LFM 3-VL. InternVL is the higher score of thinking/nonthinking per benchmark and retains a source-specific configuration identity. GPT-5.6 Luna medium; Gemini 3.5 Flash Lite minimal. Video is 2FPS/max 384 frames for models lacking native video. GPT-6 Luna grades judge-based tasks. Native published values normalized 0-100 (MME from 0-2800; OCRBench 0-1000; DocVQA/InfographicVQA 0-1). RefCOCO mean P@1 over 8 splits; POPE F 1 pooled over 3 subsets; MMMU-Pro 10-choice variant. Metric: Published normalized score (0–100).
Metric: Published normalized score (0–100). Source: huggingface.co. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Luna (Medium) | 66.8 | #79 (GPT-5.6 Luna) |
| 2 | Qwen 3.5 4B (Thinking) | 64.1 | #514 (Qwen 3.5 4B) |
| 3 | Qwen 3.5 2B (Thinking) | 59.4 | #1114 (Qwen 3.5 2B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=lfm3-vl-2b-release-activitynet-qa-test · How It Works · Data refreshed daily, snapshot 2026-11-05.