Liquid AI Releases New Visual-Language Model LFM2.5-VL-3B
Liquid AI has introduced a new visual-language model, LFM2.5-VL-3B, which shows impressive results compared to larger models.
Liquid AI has released a lightweight and fast visual-language model, LFM2.5-VL-3B.
This visual model has only 3 billion parameters, outperforming models that are more than twice its size and running directly on the Galaxy S26 Ultra at a speed of about 20 tokens per second.
LFM2.5-VL-3B can work with mobile, web, and desktop interfaces, read screens, documents, and real-world images, and accurately bind objects to coordinates.
On ScreenSpot-v2, the model scores 80.7 compared to 51.2 for Gemma-4-E4B.
It also surpasses significantly larger models in tasks related to interfaces, object binding, and text recognition.
On the M5 Max, the speed reaches 228 tokens per second, requiring only about 3 GB of memory to operate.
The model is already available on Hugging Face: https://huggingface.co/LiquidAI/LFM2.5-VL-3B
Why it matters
AnalysisThis model demonstrates significant progress in visual-language technologies, opening new possibilities for interactive applications. Its efficiency on mobile devices makes the technology more accessible to a wider audience.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
GPT Image Generative Models Prompting Guide
OpenAI's gpt-image generative models are designed to produce production-quality visuals and fully controllable creative workflows. They are suitable for both professional design ta...
Claude Sonnet 4.6 vs GPT-5.5 vs Gemini 3.1 Pro: 2026 model comparison
Every few weeks, a new model comes out, and every time the same questions arise: should I switch? what changed what to pay for?
MIT researchers show GPT-6 Astra can improve other AI models by editing their weights
In an experiment, Astra raised Qwen3.5-0.8B's accuracy on MATH-500 from 7.9% to 48.1% by changing only 0.016% of the model's weights.
GPT-6 Astra reportedly edits another AI model's weights without retraining
GPT-6 Astra is claimed to modify only 0.016% of Qwen3.5-0.8B's weights without gradients or retraining, improving reasoning, math results, and instruction following.