Liquid AI · 2026-08-12 · major
LFM2.5-VL-3B — Liquid AI's 3B vision model reads screens on a phone
LFM2.5-VL-3B is Liquid AI's new 3.1B open-weights vision-language model for phones and laptops. Screen understanding reaches 78.7 on ScreenSpot-v2 Desktop, and object grounding rises from 57.1 to 87.9 on RefCOCO-avg.
A 3.1B open-weights vision model that reads phone and desktop screens, points at objects, and calls tools on-device.
Quick facts
| Maker | Liquid AI |
|---|---|
| Parameters | 3.1B |
| License | lfm1.0 |
| Vision encoder | SigLIP2 400M NaFlex |
| Memory | About 3 GB |
| Availability | Open weights on Hugging Face |
| What's new | Screen reading, grounding, multi-image, tool calls |
Benchmarks
What is it?
Screen and interface reading is the headline change in LFM2.5-VL-3B, Liquid AI's 3.1B open-weights vision-language model. The model handles mobile, web, and desktop screens, grounds objects to coordinates, reads text and charts, and calls tools from either a text or an image prompt. Weights are on Hugging Face with no waitlist.
How does it work?
A SigLIP2 400M NaFlex vision encoder from Google sits in front of the same pre-trained backbone Liquid AI used for its LFM2.5-2.6B text model. Pre-training ran on roughly 34 trillion tokens with four times more vision data than the previous version, and the vocabulary doubled to 128,000 tokens. There is no reasoning step — the model answers directly, which keeps latency low.
Why does it matter?
Screen-reading agents have mostly needed a cloud model, which means shipping screenshots to a server. LFM2.5-VL-3B fits in about 3 GB of memory, so it runs on a laptop or a phone and no image leaves the device. Liquid AI measures 228 tokens per second on an Apple M5 Max, 116 on an AMD Ryzen AI Max+ 395, and 20 on a Galaxy S26 Ultra.
Who is it for?
on-device and edge AI developers
Frequently asked questions
- Can LFM2.5-VL-3B actually run on a phone?
- Liquid AI reports 20 tokens per second for LFM2.5-VL-3B on a Galaxy S26 Ultra, with the model fitting in roughly 3 GB of memory. That is slow next to a laptop but usable for short screen-reading and grounding tasks. Because inference stays local, no screenshot or photo is sent to a server.
- How much better is LFM2.5-VL-3B than LFM2-VL-3B?
- The biggest gains for LFM2.5-VL-3B over LFM2-VL-3B are in grounding and tool use. RefCOCO-avg goes from 57.1 to 87.9, and ToolSandbox roughly doubles from 26.4 to 59.5. Desktop screen understanding on ScreenSpot-v2 reaches 78.7, where the previous version scored 6.0.
- Does LFM2.5-VL-3B think step by step before answering?
- No. Liquid AI built LFM2.5-VL-3B as a non-reasoning model that replies directly instead of producing a chain of thought first. The trade-off is deliberate: skipping the reasoning pass keeps latency low enough for real-time and on-device use, at the cost of the extra accuracy a reasoning model can buy on hard problems.
- How fast is LFM2.5-VL-3B on a server GPU?
- On an H100, Liquid AI measures about 11,000 output tokens per second for LFM2.5-VL-3B at high concurrency. So while the model is built for edge hardware, it also batches well enough to serve many users from one GPU, which suits high-volume document or screenshot processing.
Try it
https://huggingface.co/LiquidAI/LFM2.5-VL-3B