Summary
On August 12, 2026 Liquid AI released LFM2.5-VL-3B, a 3.1-billion-parameter open-weight vision-language model built to run on phones, laptops, and single GPUs rather than in a data center. It adds stronger screen understanding, object grounding, multi-image reasoning, and function calling, extending last year's LFM2-VL-3B.
What changed
Liquid AI published LFM2.5-VL-3B with open weights: 3.1B parameters, pre-trained on ~34T tokens with 4x more vision data than its predecessor and a vocabulary doubled to 128,000 tokens. It targets on-device inference and improves screen understanding, object grounding, multi-image reasoning, and tool calling for agent use.
Why it matters
Compact vision-language models with function calling let agents see and act on screens and images locally, without sending data to a cloud endpoint—useful for latency, cost, and privacy-sensitive on-device automation. It advances the edge-agent stack as an alternative to routing every multimodal step through hosted frontier models.
Evidence excerpt
LFM2.5-VL-3B, a 3.1-billion-parameter open-weight vision-language model built to run on phones, laptops, and single GPUs rather than in a data center, extends LFM2-VL-3B with stronger screen understanding, object grounding, multi-image reasoning, and function calling.