LiquidAI releases LFM2.5-VL-3B, a 3.1B-parameter vision-language model optimized for edge devices

LiquidAI released LFM2.5-VL-3B, a 3.1B-parameter vision-language model optimized for edge and on-device deployment. The model introduces four key improvements over prior releases: enhanced screen and UI understanding, improved grounding and object detection with natural language queries, better multi-image reasoning, and significantly stronger function-calling capabilities in both text-only and vision-text contexts.

The model pairs a SigLIP2 400M NaFlex vision encoder with the same pre-trained backbone as LiquidAI’s LFM2.5-2.6B text model. Training involved approximately 34T tokens, with 4x more vision data than previous versions, drawn from curated and synthetic datasets covering image-caption pairs, OCR, grounding, and instruction-following tasks. To support non-Latin scripts, the tokenizer’s vocabulary was expanded to 128K tokens without retraining from scratch.

Benchmark results across 20 vision and multimodal benchmarks show LFM2.5-VL-3B leading its size class on real-world image tasks, including multilingual visual comprehension, instruction following, document understanding, and screen/UI comprehension. On text-only tasks, the model demonstrates competitive instruction-following and tool-use performance, matching or exceeding models up to twice its size in some evaluations.

LFM2.5-VL-3B achieves up to 228 tokens per second on an M5 Max and 116 tokens per second on a Ryzen AI Max+ 395, with a memory footprint of about 3 GB. On a Galaxy S25 Ultra, it reaches 20 tokens per second, enabling fully on-device operation. On GPU, the model delivers high throughput, reaching approximately 11K tokens per second at high concurrency, roughly twice the speed of larger 4B-class models.

The model ships with day-one support across multiple inference backends, including llama.cpp, MLX, vLLM, SGLang, and ONNX. LiquidAI provides installation instructions via the Transformers library, with compatibility for versions 5.10.1 and higher.