Liquid AI shipped LFM2.5-VL-3B this week, a 3.1B-parameter vision-language model aimed squarely at running on the device in your hand rather than in someone’s datacenter. The headline numbers are good. The one that actually changes what you can build with it is buried further down the sheet.
Grounding is the whole game
RefCOCO grounding goes from 57.1 to 87.9. That is not a polish pass, that is the difference between a model that can tell you there’s a submit button somewhere on screen and one that can hand you the coordinates. ScreenSpot-v2 averages 80.7, which is the UI-specific version of the same capability: point at the thing, not describe the thing.
Every mobile agent demo of the last two years has run into the same wall. The model reads the screenshot fine, narrates what it sees fine, and then produces a tap location that lands twenty pixels off and opens the wrong menu. Captioning benchmarks never caught this because captioning doesn’t require spatial precision. Grounding does.
Add function calling, which is new to the VL line here, and the loop closes. ToolSandbox moves from 26.4 to 59.5. Look at a screen, locate an element, call a tool with the result. Previously that meant stitching together a vision model, a separate tool-calling model, and glue code that reconciled two different notions of what was on screen. Now it’s one 3 GB artifact.
The size number matters more than the speed number
228 tokens/s on an M5 Max is a fine figure, and it’s also the least interesting one, because an M5 Max is a laptop with a fan and a power budget. The 3 GB footprint is the constraint that decides whether this ships inside an app or stays a research demo. Three gigabytes is loadable on a phone with 8 GB of RAM without evicting everything else. Whether it holds that throughput on a mid-range Android chip is a separate question, and the sources don’t answer it.
I’d also treat ToolSandbox at 59.5 as a starting position rather than a solved problem. Roughly two out of five tool interactions still go wrong. For a UI automation agent, a 40% failure rate on the action step compounds fast across a multi-step task, and nobody’s shipping unattended phone automation on those odds. Assisted flows where a human confirms each action, though, are perfectly viable at that number.
What this pressures
The cloud VLM APIs have been coasting on the assumption that anything genuinely useful needs to phone home. A 3B model that grounds at 87.9 and calls tools locally undercuts that on latency, on cost, and on the privacy pitch that matters for anything touching a screen full of personal data. Screenshots are among the most sensitive things a user can hand a model, and not sending them anywhere is a real feature.
What I want to see next is someone running this against a live app for a hundred consecutive steps and reporting where it drifts. Benchmarks are single-shot. Agents are not. Liquid AI has the weights on Hugging Face, so that test is available to anyone willing to run it.