LiquidAI Launches LFM2.5-VL-3B: Lightweight VLM Runs Tool Calling Locally
helloiamleonie · x · 2026-08-13
LiquidAI has officially released LFM2.5-VL-3B, a lightweight vision-language model. Built on the LFM2.5-2.6B base with a SigLIP2 vision encoder, it can read screens, documents, and the physical world, supporting tool calling directly from text or image inputs.
Performance Highlights:
- Scores 80.7 on ScreenSpot-v2, beating Gemma-4-E4B.
- Achieves 73.1 on RealWorldQA, outperforming InternVL-3.5-4B.
A developer tested the model locally on an M3 Pro 18GB device. It successfully recognized a photo of an English breakfast and executed a web search for the recipe, demonstrating strong capabilities in UI understanding, OCR, and tool calling.
Related event: LiquidAI Releases LFM2.5-VL-3B Edge Vision-Language Model(10 posts)→
More from Models
- Grok 4.6 Model Card Analysis: Big Internal Gains, Lags on Public SWE Evals — scaling01 · 2026-08-13
- Anthropic's Leaked Deck Predicts 2026 as the Last Window to Catch Up in AI — imjustnewatai · 2026-08-13
- Researcher Criticizes Frontier Models for Hiding Reasoning Traces, Calls for Open AI Science — rao2z · 2026-08-13
- AI Safety Researcher Points Out Typos and Rushed Third-Party References in Day-1 Model Cards — Miles_Brundage · 2026-08-13
- SemiAnalysis Slams NVIDIA: Committee-Based Frontier Model Development Does Not Work — teortaxesTex · 2026-08-13
- Elon Musk Hints at Grok 4.6, Claiming 'Pareto Gold' Dominance — shaunmmaguire · 2026-08-13