LFM2 Mixture-of-Experts Liquid Model High-Speed CPU Inference

📌 Background

Founded in 2023 by researchers from MIT CSAIL, Liquid AI replaces the traditional Transformer architecture with Liquid Neural Networks (LNN). In July 2025, the company launched the LFM2 hybrid architecture model, focusing on edge device deployment without cloud dependency. The LFM2-24B-A2B, released in March 2026, utilizes a MoE architecture with a 32K context window, redefining the cost-performance ratio of large models through ultra-fast Decode and Prefill efficiency on CPUs.

👤 Target Customers

Enterprise clients requiring localized, low-latency inference, including industrial real-time control, edge device manufacturers, privacy-sensitive industries (healthcare, finance), and SMEs unable to rely on cloud infrastructure.

💰 Revenue Streams

1) Billing based on API usage, or licensing fees for private deployment, along with model fine-tuning service fees and annual enterprise-level technical support fees; 2) Excess capacity: tiered overage charges for usage exceeding plan limits, with additional fees for dedicated capacity; 3) Proprietary implementation: fees for localized deployment and integration with internal systems, charged per project.

🧮 Cost Structure

Computational costs for model R&D and training, core technical team compensation, developer ecosystem and open-source community operations, and personnel for customer deployment and technical support.

🛡️ Moat

Originality of the MIT-backed Liquid Neural Network technology, technical barriers surrounding the non-Transformer dynamic weight update architecture, and the differentiated capability of high-speed MoE inference on CPUs.

🔑 Keys to Success

• Continuously strengthen CPU inference performance barriers to widen the gap with traditional architectures
• Rapidly expand developer toolchains and application templates to lower entry barriers
• Deepen strategic partnerships with chip manufacturers (e.g., AMD) for hardware optimization

⚠️ Risks

• The performance ceiling of Liquid Neural Networks on complex tasks has not yet been fully validated
• Open-source alternatives may dilute the uniqueness of the technology

🏢 Cases

• LFM2-24B-A2B is available for download and commercial use on Hugging Face
• LFM2.5-2.6B is already capable of running on Raspberry Pi and mobile devices

📊 SWOT Analysis