Liquid AI recently launched LFM2.5, a next-generation family of small foundation models based on the LFM2 architecture, focusing on edge devices and local deployment. This model family includes LFM2.5-1.2B-Base and LFM2.5-1.2B-Instruct, and also expands to variants in Japanese, vision-language, and audio-language. These models are released with open-source weights on Hugging Face and are showcased on the LEAP platform.

LFM2.5 retains the hybrid LFM2 architecture designed for CPU and NPU, aiming to achieve fast and memory-efficient inference. The pre-training phase of this model expanded parameters to 120 million, and training data increased from 10 trillion tokens to 28 trillion tokens. Subsequently, the instruction variant model underwent supervised fine-tuning, preference alignment, and large-scale multi-stage reinforcement learning, focusing on instruction following, tool usage, math, and knowledge reasoning.

image.png

In terms of text model performance, LFM2.5-1.2B-Instruct is the main general-purpose text model. The Liquid AI team reported performance on multiple benchmarks such as GPQA, MMLU Pro, IFEval, and IFBench, achieving scores of 38.89 on GPQA and 44.35 on MMLU Pro. These scores are significantly higher than other similar open-source models, such as Llama-3.2-1B Instruct and Gemma-3-1B IT.

Additionally, LFM2.5-1.2B-JP is a text model specifically optimized for Japanese, tailored for tasks such as JMMLU, M-IFEval, and GSM8K in Japanese. This checkpoint outperforms general instruction models on Japanese tasks and competes with other small multilingual models in these local benchmark tests.

Regarding multimodal edge workloads, LFM2.5-VL-1.6B is the updated vision-language model in this series, incorporating a visual module for image understanding. This model is optimized to support practical applications such as document understanding, user interface reading, and multi-image reasoning, and can efficiently run in edge environments.

LFM2.5-Audio-1.5B is a native audio language model that supports text and audio input and output, using a new audio tokenizer that is eight times faster than previous solutions, suitable for tasks such as real-time voice-to-voice conversation agents and automatic speech recognition.

https://www.liquid.ai/blog/introducing-lfm2-5-the-next-generation-of-on-device-ai

Key points:   

🌟 LFM2.5 is a family of small foundation models based on the LFM2 architecture, supporting various variants including text, vision-language, and audio-language.  

📈 This model performs excellently on multiple benchmarks, especially surpassing similar models on GPQA and MMLU Pro.  

🌐 The LFM2.5 series covers multimodal and regional optimization, providing strong edge computing capabilities, suitable for various practical application scenarios.