Liquid AI releases LFM2.5-VL-3B edge vision model
Liquid AI's new 3.1B-parameter vision-language model nearly matches 4.7B-parameter rivals across 28 benchmarks, with large gains in screen understanding, grounding and tool use for on-device agent workloads.
Liquid AI has released LFM2.5-VL-3B, which it calls its most capable vision-language model. The 3.1B-parameter model is available on Hugging Face and the company's Playground. It is a non-reasoning model: it answers directly, without a reasoning step, to keep latency low for real-time and on-device use.
The biggest changes from LFM2-VL-3B are in GUI and grounding tasks. ScreenSpot-v2 scores rose from single digits to an average of 80.7. That beats Qwen3.5-4B (78.5) and Gemma-4-E4B (51.2) but trails InternVL-3.5-4B (84.1). RefCOCO precision@1 rose from 57.1 to 87.9. On multi-image tasks, BLINK went from 50.2 to 61.5 and MUIRBench from 34.9 to 58.3. Function calling is new to the VL line: ToolSandbox rose from 26.4 to 59.5 and BFCL v4 from 20.5 to 32.5.
Across 28 vision benchmarks, the model averages 69.4. That ties InternVL 3.5 4B and is 0.7 points behind Qwen3.5-4B (70.1). It still trails larger models on some document and STEM tests, including DocVQA, InfographicVQA, OCRBench v2 and MMMU-Pro.
The model shares a pre-trained base with Liquid's LFM2.5-2.6B text model and uses a SigLIP2 400M NaFlex encoder. Liquid says it pre-trained on about 34T tokens, doubled the vocabulary to 128K to better support non-Latin scripts, and used 4x more vision pretraining tokens. Post-training combined supervised fine-tuning, distillation from a larger teacher model and multi-reward reinforcement learning.
Liquid also says the model runs faster on CPUs and GPUs than competitors, including smaller ones, but the material reviewed did not include latency figures. All benchmark results are the company's own.