LFM2.5-VL-3B model released with improved screen reading and function calling for edge devices

The vision-language model LFM2.5-VL-3B has been released with 3.1 billion parameters, designed for deployment on edge devices. According to the developers, it brings four main improvements over the previous version: understanding screens and user interfaces on different types of devices, more accurate grounding and object detection based on natural language, better reasoning over multiple images simultaneously and significantly stronger function calling in both text and visual-text tasks.

The model combines the SigLIP2 visual encoder with 400 million parameters (the NaFlex variant) with the LFM2.5-2.6B language backbone. Pretraining used approximately 34 trillion tokens, with four times the volume of visual data compared with the previous version. The tokenizer vocabulary was expanded to 128K entries to support languages that use non-Latin scripts. Post-training took place in two phases: supervised fine-tuning with knowledge distillation from a larger teacher model and so-called Antidoom training, followed by reinforcement learning with multiple reward signals.

According to the published results, the model achieves an average score of 69.4 across a set of visual benchmarks (including MMStar, ChartQA, DocVQA, RefCOCO, BLINK, ScreenSpot), matching the result of the larger model InternVL 3.5 4B, while Qwen3.5-4B (also a larger model) achieved a higher average of 70.1. In text tasks focused on tool use and function calling, LFM2.5-VL-3B is comparable to Gemma-4-E2B and Qwen3.5-2B. The model has had support in llama.cpp, MLX, vLLM, SGLang and ONNX since release. According to the published measurements, it handles 228 tokens per second on the M5 Max chip, 116 tokens per second on Ryzen AI Max+ 395, fits into roughly 3 GB of memory and runs at 20 tokens per second on the Galaxy S26 Ultra phone.

You can find details in the source article.