Liquid AI announced the LFM2.5-VL-3B, a vision-language model with 3.1 billion parameters aimed at edge devices. The company says the release expands capabilities in screen and interface understanding, grounding, multi-image input, and function calling.
Training
The model combines the SigLIP2 visual encoder with 400 million parameters with the pre-trained base of LFM2.5-2.6B. According to the company, around 34 trillion tokens were used in pre-training, with four times more vision data than in previous versions, including caption datasets, OCR, grounding and instructions.
The tokenizer was extended to double the vocabulary to 128,000 tokens, aiming to improve support for non-Latin alphabets, without retraining the model from scratch. In post-training, the company combined supervised fine-tuning with knowledge distillation, Antidoom training and reinforcement learning with multiple rewards.
Performance
In vision benchmarks, the LFM2.5-VL-3B achieved an average of 69.4 points, versus 57.2 for the previous version. Liquid AI says the model leads its class in tasks with real images and stands out in reading documents, charts and interface elements.
In text-only tests, the company reported improvement in tool use, with performance equivalent to Gemma-4-E2B and Qwen3.5-2B in function calling.
Inference speed
The company reported that the model runs entirely on-device at 228 tokens per second on an M5 Max, 116 tokens per second on a Ryzen AI Max+ 395 and 20 tokens per second on a Galaxy S26 Ultra, using about 3 GB of memory.
On GPUs, the LFM2.5-VL-3B reaches approximately 11,000 tokens per second in high-concurrency scenarios, roughly double that of 4B-class models, and can generate nearly 1 billion tokens per day on an H100. The model has launch compatibility with llama.cpp, MLX, vLLM, SGLang and ONNX.
Availability
The LFM2.5-VL-3B can be loaded with the Transformers library, version 5.10.1 or higher, from the identifier LiquidAI/LFM2.5-VL-3B. Liquid AI also made available a browser demo that allows sending images and testing grounding, OCR and tool use.
