Liquid AI Releases Ultra-Compact Language Model

Liquid AI, founded by former MIT researchers, has released LFM2.5-230M, the company's smallest language model. With just 230 million parameters—a metric indicating model complexity and scale—it is designed to run on devices such as smartphones, laptops, and robots without requiring network connectivity.

In the competitive landscape of AI model development, major companies like OpenAI and Google typically compete with models containing hundreds of billions to trillions of parameters. In contrast, the "edge AI" domain, which does not rely on cloud infrastructure, prioritizes achieving practical performance with minimal resources. LFM2.5-230M is positioned as a model that embodies this direction.

According to Liquid AI, the model outperforms models with more than four times its parameters on certain benchmarks. Specifically, in data extraction tasks, it achieved higher scores than Alibaba's Qwen3.5-0.8B (Instruct) with 800 million parameters and Google's Gemma 3 1B with 1 billion parameters. The company also disclosed that the model was pretrained using 19 trillion tokens of text data.

Technically, it employs a proprietary LFM2 architecture that differs from the conventional Transformer. By combining convolutional processing optimized for short-range information and attention mechanisms for broad contextual reference, it achieves fast inference while reducing memory consumption. Memory usage remains below 400MB, and on the Samsung Galaxy S25 Ultra, it achieves a processing speed of 213 tokens per second. Another notable feature is its ability to handle up to 32,000 tokens of context simultaneously.

In terms of licensing, it is offered free of charge to individuals and enterprises with annual revenue below $10 million. Companies exceeding this threshold require a paid enterprise license. Liquid AI positions this model for engineers and developers working on data extraction pipelines and autonomous edge system development.

The significance of this model lies in demonstrating that improved architecture provides an alternative to the assumption that "larger scale means greater strength." On-device AI processing offers advantages unavailable in cloud-based approaches, including privacy protection, low latency, and offline operation. Future attention will focus on how effectively this model can be deployed in sectors such as industrial robotics, medical devices, and IoT systems—environments where network connectivity is unreliable or unavailable.