Liquid AI Unveils LFM 2.5 and LFM 3 Models at CES 2026

Liquid AI's Innovative Approach

• 💡 Liquid AI focuses on building efficient generative AI models that run fast directly on any processor, distinguishing them from cloud-based solutions.
• 🚀 They develop liquid foundation models (LFMs), not traditional transformer models, with a hardware-in-the-loop approach to optimize neural architectures for specific hardware.
• ✅ The core goal is to substantially reduce the computational cost of intelligence without sacrificing quality, enabling frontier model performance on various devices.
• 🔒 These models offer key benefits like privacy, speed, and continuity, ensuring seamless operation across both online and offline workloads.

Introducing LFM 2.5

• 🎯 Liquid AI announced LFM 2.5, their most advanced tiny class of models, featuring only 1.2 billion parameters.
• 📈 This model demonstrates best-in-class instruction following capabilities, outperforming larger models like DeepSeek and Gemini Pro directly on devices.
• 🧩 LFM 2.5 includes five specialized instances: a chat model, an instruct model, a Japanese enhanced language model, a vision language model, and a lightweight audio model.
• 🛠️ These models are highly optimized for AMD Ryzen AI CPUs, GPUs, and NPUs and are available open-weight on Hugging Face and Liquid AI's Leap platform.

The Future with LFM 3

• ⚡ The company also introduced LFM 3, designed to be natively multimodal, capable of processing text, vision, and audio as input.
• 🗣️ LFM 3 will deliver audio and text outputs in 10 different languages with sub-millisecond latency for audiovisual data.
• 🗓️ This advanced multimodal model is anticipated to be released later in the year, further expanding on-device AI capabilities.

Enabling Proactive AI Agents

• 🧠 Liquid AI's models facilitate a shift from reactive AI assistants to proactive agents that run continuously and rapidly on devices.
• 🤖 A demo showcased a proactive agent capable of joining meetings, analyzing emails, and drafting responses, all performed locally and offline.
• ✅ The emphasis is on user control and the ability for these agents to perform tasks in the background, without relying on cloud inference.
• 🤝 Liquid AI is collaborating with Zoom to integrate these proactive features into their platform, enhancing user experience.

Defining the Next Phase of AI

• 🔑 Ramin Hasani highlighted that the next phase of AI adoption will be defined by efficiency, latency, and deployability, rather than just raw scale.
• 💡 Liquid AI's design philosophy is to maximize intelligence per unit of compute, ensuring high-quality AI can run locally and in real-time across various hardware.
• 🌐 This approach aims to make high-quality AI accessible and functional directly on devices, from phones to airplanes.