Liquid AI's LFM2.5-VL-3B Can Run on Your Phone and Still Read Your Screen
The new 3-billion-parameter vision model fits in 3 GB of memory, beats larger rivals on screen-reading benchmarks, and runs fast enough on a smartphone to be genuinely useful.
Key points
- Liquid AI released LFM2.5-VL-3B on Hugging Face, a vision-language model small enough to run on a laptop or smartphone without a cloud connection.
- The model scores 78.7 on the ScreenSpot-v2 Desktop benchmark for reading digital screens, up from 6.0 for its predecessor LFM2-VL-3B.
- On a single Nvidia H100 GPU, the model processes roughly one billion output tokens per day, about twice as fast as comparable 4-billion-parameter models.
- The model fits in approximately 3 GB of memory and reaches 20 tokens per second on a Samsung Galaxy S26 Ultra smartphone.
- LFM2.5-VL-3B adds function calling, a feature that lets the model trigger actions in apps and tools, to its previous image-understanding abilities.
Liquid AI has released LFM2.5-VL-3B, a vision-language model (software that understands both images and text, like a more capable version of the image-description tools built into some phones) designed to run directly on everyday hardware. No internet connection to a data centre required.
The headline number is size. At 3.1 billion parameters (the internal values that shape how a model thinks), it fits in about 3 GB of memory. That is smaller than many mobile games.
What can it actually do?
The model reads documents, decodes text in images, identifies where objects sit in a photo, and understands what is on a computer or phone screen. On the ScreenSpot-v2 benchmark, a standard test for reading on-screen interfaces, LFM2.5-VL-3B scored 78.7 for desktop, 81.2 for mobile, and 82.2 for web. Its predecessor, LFM2-VL-3B, scored 6.0, 7.6, and 2.5 on the same three tests.
The model also supports function calling: the ability to trigger actions in other software, such as clicking a button or filling a form, based on what it sees on screen. Liquid AI says performance here is now on a par with Google's Gemma-4-E2B and Alibaba's Qwen3.5-2B, two competing small models.
| Model | Size | ScreenSpot-v2 Desktop | MathVista (mini) | Avg vision score |
|---|---|---|---|---|
| LFM2.5-VL-3B | 3.1B | 78.7 | 68.5 | 69.4 |
| LFM2-VL-3B | 3.1B | 6.0 | 62.1 | 57.2 |
| InternVL 3.5 4B | 4.7B | 82.0 | 67.1 | 69.4 |
| Qwen3.5-4B | 4.7B | 76.3 | 63.6 | 70.1 |
| Gemma-4-E4B-it | 8B | 45.8 | 45.2 | 59.7 |
How fast does it run?
Fast enough to matter on real devices. On Apple's M5 Max chip (the kind in a high-end MacBook Pro), it processes 228 tokens per second, where a token is roughly three-quarters of a word. On AMD's Ryzen AI Max+ 395 (a chip found in certain Windows laptops), it reaches 116 tokens per second. On a Galaxy S26 Ultra smartphone, it hits 20 tokens per second, which is quick enough for a live conversation.
On a server-grade Nvidia H100 GPU (the specialised chip that does heavy AI number-crunching in data centres), it reaches around 11,000 tokens per second at high load, nearly twice as fast as 4-billion-parameter rivals.
Who is this for?
Developers building apps that need to understand images or screens without sending data to a remote server. Think accessibility tools that describe what is on screen, expense apps that read receipts, or business software that fills forms automatically. The model is available now on Hugging Face, the platform for sharing AI models, under Liquid AI's account.
