Liquid AI Releases LFM2.5-VL-3B for On-Device Vision Tasks
The 3.1-billion-parameter model adds screen understanding, multi-image input, grounding and function calling, while its speed and benchmark claims remain first-party results.
Liquid AI released LFM2.5-VL-3B on August 12, expanding its compact model family into a 3.1-billion-parameter vision-language system designed for local and edge inference. The company's release makes the model available on Hugging Face and focuses the update on screen understanding, object grounding, multi-image input and function calling.
The practical change is not simply another small image model. Liquid AI is packaging visual interpretation and tool use for devices where memory, latency or data handling make a remote model less attractive. Developers working on document interfaces, visual assistants or device-control workflows are the clearest audience, although the published performance record comes from Liquid AI itself.
What changed in the multimodal stack
LFM2.5-VL-3B combines a 400-million-parameter SigLIP2 NaFlex vision encoder with the same pre-trained backbone used in Liquid AI's 2.6-billion-parameter text model. The company says it pre-trained the combined system on about 34 trillion tokens, used four times more vision data than its previous version and doubled the vocabulary to 128,000 tokens to improve non-Latin-script coverage.
The four stated capability priorities are specific. Screen and user-interface understanding targets content displayed across devices. Grounding connects natural-language requests to objects or regions in an image. Multi-image input lets the model reason across more than one visual frame. Function calling extends tool use to text-only and mixed vision-text interactions. Those are useful product capabilities, but their reliability in a particular application still depends on the application's images, prompts and tools.
What on-device use looks like
Liquid AI reports that the model fits in roughly 3 GB of memory. In its tests, decoding reached 228 tokens per second on an Apple M5 Max, 116 on a Ryzen AI Max+ 395 and 20 on a Galaxy S26 Ultra. The company also reports higher-throughput server results at concurrency, including about 11,000 output tokens per second in its comparison. These are vendor measurements, not portable guarantees for every quantisation, input size, runtime or thermal envelope.
Day-one runtime support spans llama.cpp, MLX, vLLM, SGLang and ONNX. The release also includes a browser-based WebGPU demonstration and points developers to fine-tuning material. That breadth matters because a model described as on-device is only useful when teams can run it through the deployment stack they already support.
Evidence quality and limits
Liquid AI evaluated the model across vision, instruction-following and tool-use benchmarks with vLLM 0.26.0 and recommended generation settings where available. Its tables compare several models of different effective sizes and architectures. The company says LFM2.5-VL-3B leads its size class on real-world image tasks, but those rankings have not been independently reproduced in the evidence reviewed for this article.
The same caution applies to speed. A phone result shows that local execution is possible; it does not establish acceptable latency, battery impact or accuracy for a production camera workflow. The source also does not provide field testing across the variety of screens, documents and languages implied by the capability list. Teams should test the exact model build, runtime and task distribution they intend to ship.
Availability and next checkpoint
The model is downloadable from Hugging Face now, and the browser demo exposes image upload, grounding, optical-character recognition and tool-use examples. Availability is therefore verifiable even though the comparative quality and speed claims remain first-party evidence.
Status
Confirmed. Liquid AI has released the model and public usage path. Internal confidence is medium because the capability, benchmark and inference measurements come from the model developer without independent replication in the fetched evidence.
Sources
Update note: Last reviewed 2026-08-17. We will revise this post if independent evaluations, revised model artifacts or measured device limits materially change the picture.
Sources
- Liquid AI team article — LFM2.5-VL-3B — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.