Liquid AI Releases LFM2.5-VL-3B — An Open Vision-Language Model for...

Liquid AI Releases LFM2.5-VL-3B — An Open Vision-Language Model for Devices

On August 12, 2026, Liquid AI released LFM2.5-VL-3B, an open vision-language model with 3.1B parameters that scores an average of 69.4 points across 28 vision benchmarks, compared to 57.2 for its predecessor, and runs in ~3 GB of memory.

Compiled by Sergey KostenchukPublished 2026-08-15Updated 2026-09-18

2026-08-15 Research Meta

🤖 Liquid AI Releases LFM2.5-VL-3B — An Open Vision-Language Model for Devices

The model is built on the LFM2.5-2.6B text base and the SigLIP2 400M NaFlex image encoder: 32,768-token context, 16 languages including Russian, ~3 GB of memory, and responses without a chain of thought.

🌍 This is a new reference for on-device VLMs: the 3B class for the first time shows the level of 4B models in screen understanding (ScreenSpot-v2 80.7) and function calling (ToolSandbox 59.5) — useful for local UI agents working with interfaces without the cloud.

👤 You can try it right now: a WebGPU demo in the browser, the Liquid AI Playground, and GGUF, ONNX, and MLX builds for llama.cpp, LM Studio, Ollama, and Jan. On the Galaxy S26 Ultra, the model maintains ~20 tokens/sec.

Source 1: https://www.liquid.ai/blog/lfm2-5-vl-3b Source 2: https://huggingface.co/LiquidAI/LFM2.5-VL-3B

Sources