Meet LFM2

Liquid.ai's new Liquid Foundation Models are redefining on-device AI. Discover how this novel architecture delivers unparalleled speed, efficiency, and privacy for the next generation of edge applications.

⚡️ Blazing Fast

Up to 2x faster decode and prefill speeds on CPU compared to leading competitors.

🧠 Highly Efficient

A 3x improvement in training efficiency makes powerful AI more cost-effective than ever.

🔒 Private by Design

Runs entirely on-device, ensuring millisecond latency and data-sovereign privacy.

A Novel Hybrid Architecture

LFM2 moves beyond traditional transformers, blending the strengths of convolutions and attention. This section breaks down its unique 16-block structure, the core of its efficiency. Hover over the blocks to learn more.

Start
→
Conv Block (x10)
Double-Gated Short-Range LIV Convolution: Optimized for embedded CPUs, these blocks use dynamic, input-aware weights for maximum efficiency.
+
GQA Block (x6)
Grouped Query Attention: A more efficient form of attention, combined with SwiGLU activation and RMSNorm for powerful, stable processing.
→
Output

This 16-block hybrid structure was discovered by STAR, Liquid AI's neural architecture search engine, which optimized for real-world performance on hardware like Qualcomm Snapdragon SoCs.

Performance Dashboard

LFM2 doesn't just promise efficiency—it delivers superior performance. This interactive chart compares LFM2 models against competitors on key industry benchmarks. Use the buttons to filter by capability.

How It Works & What It Can Do

Explore the sophisticated training pipeline that gives LFM2 its power and the advanced capabilities it unlocks, from building reliable AI agents to its detailed model specifications.

LFM Open License v1.0 Explained

LFM2 is released under a "source-available with commercial restrictions" license. It's designed to be open for research and small businesses, while ensuring commercial success contributes back to the project.

✅ Permitted Free Use

  • 🔬**Research & Non-Profit:** Unrestricted use for academic and non-commercial projects.
  • 🚀**Startups & Small Businesses:** Free commercial use for companies with less than $10 million USD in annual revenue.
  • 🔧**Proprietary Fine-Tuning:** No "Copyleft" requirement. You can fine-tune the models on your own data and keep the derivatives proprietary.

⚠️ Action Required

  • 💼**Large Enterprises:** Companies with annual revenue exceeding $10 million USD must contact Liquid AI to purchase a commercial license.
  • ©️**Attribution:** You must retain all original copyright and license notices when distributing the model or its derivatives.