Editorial illustration for Liquid AI's 3B Vision Model Shows Major Gains in Screen Reading, Object Grounding
Liquid AI's 3B Vision Model Excels at Screen Reading
Liquid AI put out LFM2.5-VL-3B yesterday, a 3.1-billion-parameter vision-language model built to run on phones, laptops and other devices rather than in a data center. It reads screens across mobile, web and desktop interfaces, ties objects to specific coordinates, parses documents and charts, and calls tools from either text or image input. On Liquid AI's own tally of 28 vision benchmarks, the model averages 69.4, matching InternVL-3.5-4B and trailing Qwen3.5-4B by 0.7 points, despite both competitors carrying more parameters at 4.7B.
It skips the reasoning-chain approach in favor of direct answers, which keeps latency down. Liquid AI says the model runs in about 3 GB of memory and decodes 228 tokens per second on an Apple M5 Max.
The company shipped the checkpoint in four formats, native, GGUF, ONNX and MLX, with support from llama.cpp, MLX, vLLM, SGLang and ONNX runtimes on day one. That range matters for who can actually deploy this outside a lab. It also raises the question of who gets to use it for free, and where Liquid AI draws the line on commercial licensing.
Liquid AI reports an average of 69.4 across 28 vision benchmarks. That matches InternVL-3.5-4B and sits 0.7 points behind Qwen3.5-4B, both 4.7B models.
Why this matters
The gap between LFM2.5-VL-3B and its 4.7B rivals is small enough to matter for anyone deciding what actually ships on a phone or a laptop. A 69.4 average against InternVL-3.5-4B's identical score, with Qwen3.5-4B just 0.7 points ahead, tells us parameter count is no longer the deciding factor for screen-reading and grounding tasks. That's a real signal for developers building agents that need to click buttons, read charts, or parse a UI locally instead of round-tripping to a server.
The multi-image jumps, BLINK from 50.2 to 61.5 and MuirBench from 34.9 to 58.3, suggest Liquid AI tuned specifically for the multi-frame reasoning that on-device assistants actually need, not just single-image benchmark scores. We'd temper enthusiasm on one point: this is a non-reasoning model, so it trades chain-of-thought depth for speed and lower memory use. For founders weighing on-device AI against cloud inference costs, that trade-off is worth testing directly against your workload before assuming smaller and faster wins outright.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives OpenAI Slows AI Model Releases, Details Safety Shortcomings
From Archives Claude Agents Sabotaged Each Other on Shared Server, Hid Actions From Users
Common Questions Answered
What are the key capabilities of Liquid AI's LFM2.5-VL-3B model?
LFM2.5-VL-3B is designed to read screens across mobile, web and desktop interfaces, tie objects to specific coordinates, parse documents and charts, and call tools from either text or image input. This 3.1-billion-parameter vision-language model is optimized to run on phones, laptops and other devices rather than requiring data center infrastructure.
How does LFM2.5-VL-3B perform compared to larger competitor models on vision benchmarks?
On Liquid AI's evaluation of 28 vision benchmarks, LFM2.5-VL-3B achieves an average score of 69.4, which matches InternVL-3.5-4B and trails Qwen3.5-4B by only 0.7 points. This is notable because both competing models have 4.7 billion parameters, making LFM2.5-VL-3B's smaller size a significant efficiency advantage.
Why does the small performance gap between LFM2.5-VL-3B and larger models matter for developers?
The minimal 0.7-point difference between LFM2.5-VL-3B and its larger 4.7B rivals demonstrates that parameter count is no longer the deciding factor for screen-reading and object-grounding tasks. This makes the smaller model practical for developers building on-device agents that need to click buttons, read charts, or parse UIs locally without sending data to remote servers.
What is the parameter size of Liquid AI's LFM2.5-VL-3B and what advantage does this provide?
LFM2.5-VL-3B contains 3.1 billion parameters, making it smaller than competing models like InternVL-3.5-4B and Qwen3.5-4B which have 4.7 billion parameters. The smaller parameter count allows the model to run efficiently on mobile devices, laptops, and other edge devices rather than requiring data center resources.
Further Reading
- Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device - MarkTechPost
- LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge - Hugging Face
- LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge - Liquid AI
- Liquid AI Ships LFM2.5-VL-3B for Faster Vision-Language AI on the Edge - UNITE.AI
- Liquid AI Open-Weights Vision Model Runs Privately on Phones, Outpaces Larger Rivals - TechTimes
