LFM2 1.2B is a compact generative language model developed by Liquid AI as part of its second-generation Liquid Foundation Models (LFM2) series. Designed specifically for efficient on-device deployment, the model is optimized for edge hardware such as smartphones, laptops, and IoT devices, offering a small memory footprint that typically requires less than 1GB when quantized. It represents a significant architectural shift from standard transformer-only models, prioritizing high throughput and low latency on CPU and NPU hardware.
The model utilizes a hybrid architecture that integrates Liquid AI’s proprietary structured operators with traditional attention mechanisms. Specifically, it consists of ten double-gated short-range LIV (Liquid Implicit Variable) convolution blocks and six grouped query attention (GQA) blocks. This design enables faster training and inference speeds compared to similarly sized transformer-based models while maintaining competitive performance across standard benchmarks.
Supporting a 32,000-token context window, LFM2 1.2B is capable of handling long-context tasks such as document analysis and multi-turn dialogue. It is a multilingual model trained on a data mixture comprising approximately 75% English, 20% multilingual text, and 5% code, providing support for languages including Arabic, Chinese, French, German, Japanese, Korean, and Spanish. Its primary capabilities include instruction following, mathematical reasoning, and tool-use applications.