LFM2

LFM2 is a cutting-edge class of Liquid Foundation Models (LFMs) developed by Liquid AI. It is designed to deliver the fastest on-device generative AI experience, making it ideal for a wide range of devices and applications. LFM2 stands out for its quality, speed, and memory efficiency, setting a new standard in the industry. Built on a hybrid architecture, LFM2 offers 2x faster decode and prefill performance compared to Qwen3 on CPU, making it a top choice for efficient AI agents.

Benefits

LFM2 offers several key advantages:

• Speed and Efficiency: LFM2 provides faster performance and better memory efficiency, making it suitable for on-device and edge use cases.
• Cost-Effective Training: The new architecture and training infrastructure deliver a 3x improvement in training efficiency over the previous LFM generation, making it a cost-effective solution for building general-purpose AI systems.
• Versatility: LFM2 is designed to balance quality, latency, and memory for specific tasks and hardware requirements, making it adaptable for various applications.
• Privacy and Resilience: By shifting large generative models from distant clouds to lean, on-device LLMs, LFM2 unlocks millisecond latency, offline resilience, and data-sovereign privacy.

Use Cases

LFM2 is ideal for a variety of applications, including:

• Consumer Electronics: Enhancing the capabilities of phones, laptops, and wearables with real-time AI processing.
• Robotics and Smart Appliances: Enabling real-time reasoning and decision-making for robots and smart devices.
• Finance and E-Commerce: Providing efficient and private AI solutions for financial and e-commerce applications.
• Education: Offering AI-driven educational tools and resources.
• Defense, Space, and Cybersecurity: Supporting critical applications in defense, space, and cybersecurity with compact, private foundation models.

Performance and Benchmarks

LFM2 outperforms similarly-sized models across various benchmarks, including knowledge, instruction following, mathematics, and multilingualism. It competes with larger models like Qwen3-1.7B and Gemma 3 1B IT, demonstrating its superior performance and efficiency.

Availability and Licensing

LFM2 models are available on Hugging Face under an open license based on Apache 2.0. This license allows for academic and research use, as well as commercial use for smaller companies (under $10m revenue). Larger companies should contact Liquid AI for a commercial license. LFM2 models are designed for on-device efficiency and can be tested and fine-tuned for specific use cases.

For custom solutions with edge deployment, interested parties can contact Liquid AI's sales team at sales@liquid.ai.