Liquid AI’s LFM2.5: Compact On-Device Models

Liquid AI has introduced LFM2.5, a new family of compact on-device foundation models designed to enhance the performance of agentic applications. These models offer improved quality, reduced latency, and support for a wider range of modalities, all within the ~1 billion parameter class. LFM2.5 builds upon the LFM2 architecture with pretraining scaled from 10 trillion to 28 trillion tokens and expanded reinforcement learning post-training, enabling better instruction following. This advancement is crucial as it allows for more efficient and versatile AI applications directly on devices, enhancing user experience and functionality.

Liquid AI’s release of LFM2.5 marks a significant advancement in the realm of on-device artificial intelligence. By introducing a family of tiny foundation models, Liquid AI aims to enhance the performance of agentic applications that operate directly on devices. These models, which fall within the 1 billion parameter class, are designed to deliver higher quality and lower latency, making them particularly suitable for applications that demand quick and reliable responses. The ability to support a broader range of modalities means these models can handle various types of data inputs, enhancing their versatility in real-world applications.

The development of LFM2.5 is rooted in Liquid AI’s device-optimized hybrid architecture, which has been scaled from 10 trillion to 28 trillion tokens during pretraining. This scaling is crucial as it allows the models to learn from a vast array of data, improving their ability to understand and generate human-like responses. The expanded reinforcement learning post-training further refines these models, enabling them to adapt better to specific tasks and environments. This combination of extensive pretraining and targeted post-training ensures that LFM2.5 models are not only powerful but also finely tuned for practical use cases.

One of the standout features of LFM2.5 is its enhanced capability for instruction following. This is particularly important in scenarios where AI models are expected to execute specific tasks based on user commands. The higher ceilings for instruction following mean that these models can understand and act on more complex instructions, making them more effective in assisting users across various domains. This capability is essential for applications ranging from personal assistants to more specialized industry-specific tools, where precise and reliable execution of tasks is critical.

The release of LFM2.5 has broader implications for the future of on-device AI technology. By focusing on smaller, more efficient models that can operate independently of cloud-based systems, Liquid AI is paving the way for more secure and privacy-conscious AI applications. Users can benefit from the reduced latency and increased reliability of on-device processing, while also enjoying the peace of mind that comes with keeping their data on their own devices. As AI continues to integrate into everyday life, developments like LFM2.5 highlight the importance of creating models that are not only powerful but also adaptable to the evolving needs of users and industries.

Read the original article here