Liquid AI LFM2.5-VL-3B Targets Fast Vision AI at the Edge

Liquid AI has released an upgraded 3B vision-language model

Liquid AI released LFM2.5-VL-3B on August 12, 2026 as its most capable vision-language model in the LFM2.5 family. The model is available through Hugging Face and Liquid AI's Playground, with local deployment support across several inference runtimes.

This is a released open-weight model, not an upcoming preview. Liquid positions it for edge and on-device workloads where screen understanding, visual grounding, tool use and latency matter.

The update targets practical visual-agent capabilities

LFM2.5-VL-3B improves four areas over Liquid's earlier LFM2-VL-3B: screen and UI understanding, function calling, grounding and multi-image input.

Screen understanding is important for agents that need to interpret application interfaces or digital workflows. Grounding lets the model identify requested objects and return their image locations. Function calling extends the model from passive visual question answering toward tool-using applications, including workflows that combine text and images. Multi-image support improves tasks that require comparing or reasoning across several visual inputs.

Liquid describes LFM2.5-VL-3B as a non-reasoning model that answers directly, prioritizing lower latency rather than exposing a long reasoning mode.

Liquid reports substantial gains over the previous model

In Liquid's evaluation, ScreenSpot-v2 averages 80.7. ToolSandbox rises from 26.4 to 59.5, while BFCL v4 improves from 20.5 to 32.5. RefCOCO average precision at one increases from 57.1 to 87.9, and multi-image benchmarks also improve substantially over the earlier LFM2-VL-3B.

These are Liquid AI-reported benchmark results using its stated evaluation setup. They should not be interpreted as universal performance rankings across every model, runtime, quantization, device or application. Real deployments should test visual accuracy, tool-call reliability, latency, memory use and failure behavior on representative inputs.

Designed to run across local hardware and runtimes

Liquid says LFM2.5 models have native support across llama.cpp, MLX and vLLM, covering hardware from Apple, AMD, Qualcomm and NVIDIA. The company's broader LFM2.5 positioning emphasizes local and edge inference rather than requiring one hosted API.

Open weights also do not remove every deployment obligation. Teams need to review the applicable model license, runtime licenses, privacy requirements and any restrictions associated with downstream applications or training data.

For edge-AI developers, the release is notable because it combines a relatively compact model size with capabilities increasingly needed by visual agents: understanding interfaces, locating objects, calling tools and using multiple images. The main question is how well those benchmark gains translate to specific phones, PCs, embedded systems and robotics workloads under real latency and memory limits.