An Open-Weight VLM Built for Local and Real-Time Use

LFM2.5-VL-3B combines vision, interface understanding, and function calling in a 3B-parameter open-weight model designed to run locally.

LFM2.5-VL-3B extends the LFM2.5 family into multimodal use cases with a 3.1B-parameter model capable of analyzing screens, documents, charts, and scenes from the physical world. It can also locate objects from an instruction, process multiple images, and call functions from either text or image input.

The model is built on the LFM2.5-2.6B base and paired with a 400M-parameter SigLIP2 NaFlex vision encoder. Liquid AI says pretraining reached roughly 34 trillion tokens, vision pretraining was scaled by 4x, and the vocabulary was expanded to 128K entries. Post-training combines SFT with knowledge distillation from a larger teacher model, Antidoom training, and multi-reward reinforcement learning.

Part of the work focuses on interaction with digital interfaces. According to Liquid AI’s evaluations, LFM2.5-VL-3B averages 80.7 on ScreenSpot-v2, compared with 51.2 for Gemma-4-E4B. On RefCOCO, the average score rises from 57.1 with LFM2-VL-3B to 87.9, while ToolSandbox improves from 26.4 to 59.5. The model also reaches 84.3 on TextVQA and 73.1 on RealWorldQA. These results come from Liquid AI’s own benchmark protocol and do not constitute an independent evaluation.

LFM2.5-VL-3B remains a non-reasoning model, designed to answer directly in order to reduce latency. Liquid AI measures 228 tokens per second on an Apple M5 Max and 116 on an AMD Ryzen AI Max+ 395, using about 3 GB of memory. On a Galaxy S26 Ultra, the company reports 20 tokens per second. On an NVIDIA H100, its tests show roughly 34 ms to first token on a multi-image input and close to 11,000 output tokens per second at high concurrency.

The model is available as open weights and supports llama.cpp, MLX, vLLM, SGLang, and ONNX from day one. Liquid AI also provides GGUF, MLX, and ONNX versions, along with a WebGPU demo that runs the model directly on-device. This setup is aimed in particular at agents that can read and interact with interfaces, document analysis, OCR, grounding, and local multimodal applications.