Liquid AI Launches 3.1 Billion-Parameter On-Device Vision Model

Liquid AI is building its LFM2.5 family for edge deployment, where models run on phones, laptops and PCs rather than relying on cloud APIs. The approach matters because local inference can reduce latency, keep sensitive data on the device and eliminate per-token charges. A compact model that can also handle long inputs, invoke tools and complete multi-step workflows could broaden the market for always-on personal and enterprise agents.

Liquid AI released the open-weight LFM2.5-2.6B on Aug. 4, 2026. The dense model contains 2.69 billion parameters, was pretrained on about 34 trillion tokens and underwent a dedicated 128K context-extension phase. The company said it can decode 30 tokens a second on a phone while using less than 2.5 GB of memory, with support for planning, tool calling and multi-step tasks. Base and post-trained checkpoints are available on Hugging Face, alongside support for llama.cpp, MLX, vLLM, SGLang and ONNX.

Liquid AI’s LFM2.5 Encoder family targets language-understanding workloads such as classification, intent routing and long-document detection. The company converted a decoder backbone into a bidirectional architecture and trained it with a masked-language-model objective, allowing the models to use context from both directions. The design is optimized for CPU-only environments, aiming to reduce computing costs and make local or edge deployment practical without dedicated GPU infrastructure.

Liquid AI has released two open-weight models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, with 230 million and 350 million parameters, respectively. Both support context windows of as many as 8,192 tokens. The company said the encoders remain fast on CPUs even at 8K context, offering a compact deployment option for text classification, semantic analysis, routing and long-form content screening.