Liquid AI released LFM2.5-VL-3B, a 3.1-billion-parameter vision-language model designed to run locally on hardware ranging from smartphones to desktop graphics processors, according to a report published by Hugging Face. The model responds directly without generating chain-of-thought reasoning steps, targeting real-time on-device applications.
Hugging Face detailed four core areas where the release expands on previous models: user interface understanding across digital displays, object detection using natural language queries, multi-image reasoning, and tool calling. On the Berkeley Function Calling Leaderboard (BFCL V4), the model achieved a score of 32.5.
Engineers paired a SigLIP2 400-million-parameter NaFlex vision encoder with the pre-trained backbone from the LFM2.5-2.6B text model. Pre-training incorporated roughly 34 trillion tokens with four times more vision data than previous releases. Liquid AI doubled the tokenizer vocabulary to 128,000 tokens to handle non-Latin scripts.
Training and post-training
Post-training followed a two-stage pipeline starting with supervised fine-tuning, knowledge distillation from a larger teacher model, and Antidoom training. The second stage used multi-reward reinforcement learning.
Hardware performance and speed
Hardware benchmarks recorded decode speeds of 228 tokens per second on an Apple M5 Max processor and 116 tokens per second on an AMD Ryzen AI Max+ 395 chip. The model fits inside roughly 3 gigabytes of memory and reaches 20 tokens per second on a Samsung Galaxy S26 Ultra smartphone.
Graphics hardware tests recorded throughput of approximately 11,000 output tokens per second under high concurrency, equivalent to nearly 1 billion output tokens daily on a single Nvidia H100 GPU. Hugging Face confirmed day-one compatibility with llama.cpp, MLX, vLLM, SGLang, and ONNX.
System requirements specify Python packages including transformers version 5.10.1 or higher for local execution. Liquid AI also made a WebGPU browser demo available for browser-based image analysis, object grounding, and function execution.
