LiquidAI LFM2.5-VL-3B: a 3.1B local VLM that beats Gemma-4 E4B — screen understanding 2.5 → 82.2 TL;DR: LiquidAI released LFM2.5-VL-3B, a 3.1B vision-language model that runs fully local (llama.cpp, MLX, vLLM, even a WebGPU demo). • Beats Gemma-4 E4B (8B) 69.4 vs 59.7; edges Qwen3.5-4B • Screen understanding: 2.5 → 82.2 on ScreenSpot-v2 Web (huge jump) • 228 tok/s on M5 Max, 20 tok/s on a Galaxy S26 Ultra • Function calling / object grounding included What I found interesting is the busine...