3B to 8B Scaling: Liquid AI's New Vision Model Runs Locally

According to the latest Beating Monitor report, Liquid AI has released the visual language model LFM2.5-VL-3B, with only 31 billion parameters, targeting local devices such as smartphones and computers. It is capable of understanding web pages, app interfaces, documents, and multiple images, and has added a new feature for invoking visual tools.

Officially, it achieved an average score of 69.4 in 28 visual benchmarks. The 8 billion parameter Gemma-4-E4B scored only 59.7, while the 5.1 billion parameter Gemma-4-E2B scored 52.0. It is on par with the 4B version of InternVL 3.5 with 47 billion parameters, scoring only 0.7 points lower than the Qwen3.5-4B.

Speed is where it truly shines. After 4-bit quantization, the model can generate 228 tokens per second on the M5 Max, occupying about 3.3GB of memory; it can also achieve 20 tokens per second on the Galaxy S26 Ultra. In the case of a single H100 high-concurrency scenario, the maximum output throughput reaches approximately 11,000 tokens per second, roughly double that of some 4B-level models.

The model's weights have been made available and can be downloaded from Hugging Face, with support for llama.cpp, MLX, vLLM, SGLang, and ONNX. It is released under the LFM Open License v1.0, allowing free commercial use for companies with annual revenue below $10 million; those above this threshold will need to purchase a commercial license.