Liquid AI Releases LFM2.5-350M: A Compact 350M Parameter Model

In the current landscape of generative AI, the scaling laws have generally dictated that more parameters equal more intelligence. However, Liquid AI is challenging this convention with the release of LFM2.5-350M. This model serves as a technical case study in intelligence density, benefiting from additional pre-training from 10 trillion to 28 trillion tokens and large-scale reinforcement learning.

The significance of LFM2.5-350M lies in its architecture and training efficiency. While most AI companies focus on frontier models, Liquid AI targets edge devices—those with limited memory and compute capabilities—demonstrating that a 350-million parameter model can outperform models more than twice its size across several evaluated benchmarks.

The core technical differentiator of LFM2.5-350M is its departure from the pure Transformer architecture. It employs a hybrid structure based on Linear Input-Varying Systems (LIVs). Traditional Transformers rely entirely on self-attention mechanisms, which suffer from quadratic scaling issues. Liquid AI addresses this by utilizing a hybrid backbone consisting of Double-Gated LIV Convolution Blocks and Grouped Query Attention (GQA) Blocks, significantly reducing memory overhead while maintaining high precision.

The LFM2.5-350M was pre-trained on 28 trillion tokens with an extremely high training-to-parameter ratio, ensuring that the model’s limited parameter count is utilized to its maximum potential. This results in high intelligence density, making it suitable for high-speed, agentic tasks rather than general-purpose reasoning.

Among its key features, the model's ability to achieve high throughput while maintaining a low memory footprint makes it ideal for local deployment. It can process up to 40.4K output tokens per second on a single NVIDIA H100 GPU, which is perfect for real-time data extraction. However, the documentation explicitly states that LFM2.5-350M is not recommended for complex coding, mathematics, or creative writing.