Release of LFM2.5-VL-3B Vision Model
Liquid AI has announced the release of LFM2.5-VL-3B, a 3.1-billion parameter vision-language model (VLM). The model, which has a file size of approximately 2GB, is designed to operate locally on mobile hardware. The release includes a corresponding model card on Hugging Face.
The new iteration demonstrates significant improvements in spatial awareness compared to its predecessor. According to unverified reports from a user on Reddit, the model’s performance on the ScreenSpot-v2 desktop benchmark increased from a score of 6 to 78.7, suggesting a heightened ability to identify the precise location of objects within an image.
Local Performance and Hardware Testing
In a practical demonstration shared on Reddit, the model was tested running locally on an iPhone 17 via the atomic.chat mobile application. During this test, the model successfully identified a “Steve” figurine from the Minecraft franchise and provided a detailed, part-by-part description of the object.
However, the testing also highlighted potential latency issues during local execution. The user reported that the model required approximately two minutes and 31 seconds to process the request. Technical commentary from the same forum suggests this delay may be attributed to slow prompt processing speeds—estimated between 2 and 16 tokens per second—rather than reasoning time, as LFM2.5-VL-3B is categorized as a non-reasoning model. Observers noted that reducing input image resolution could potentially mitigate these processing delays.
Industry Context and Technical Limitations
While the model shows progress in object localization, some community feedback suggests that small-scale vision models continue to face challenges regarding complex scene understanding. Some users noted that while these models can perform “party tricks,” such as recognizing well-known fictional characters, they may struggle to interpret complex activities within an image.
Technical discussions also highlighted existing limitations for running VLMs on mobile devices. Specifically, it was noted that when using tools like llama.cpp, the mmproj files are not easily quantifiable, which prevents the use of certain acceleration kernels that could improve performance.
As the industry moves toward more efficient, smaller-scale models for agentic tasks, questions remain regarding the trade-off between model size and “world knowledge.” While LFM2.5-VL-3B offers a lightweight solution for local deployment, the balance between parameter count and comprehensive environmental understanding remains a central focus for developers in the field.