Idea
Compact foundation models delivering fast, memory-efficient AI inference for edge devices with strong multimodal and retrieval capabilities.
Research Paper
Core Innovation
This paper introduces LFM2 (Liquid Foundations Models 2), a family of foundation models combining a hybrid backbone architecture optimized via hardware-in-the-loop search for edge constraints. It innovates with a tempered Top-K knowledge distillation, curriculum learning, and a three-stage post-training pipeline, achieving superior speed (2x) and task performance on CPUs compared to similar-sized models.
Why It Matters
Edge devices require AI models that balance performance with limited compute and memory resources. LFM2 addresses this by enabling faster, efficient inference on CPUs without sacrificing accuracy, supporting diverse tasks including vision, speech, and retrieval. This scalability and open deployment facilitate broader adoption of advanced AI in real-world, resource-constrained environments. Top use cases: Privacy-first Personal Assistants , Real-time Speech Translation, Chat w/ local files (Offline RAG), Smart Vision
Market Size (TAM)
$20–50B TAM for edge AI inference models; $5–10B SAM from mobile, IoT, and enterprise edge deployments. Driven by demand for low-latency, privacy-preserving AI and growth in edge computing.
Potential Customers & Pain Points
- Edge device manufacturers – Need efficient AI models for limited hardware
- Mobile app developers – Require fast low-latency inference
- Enterprises with on-premise AI – Demand privacy-preserving local processing
- AI researchers – Seek open versatile models for multimodal tasks
- Cloud providers – Aim to reduce inference costs and latency
Business Model
Open-source model weights and deployment packages with commercial licensing for enterprise integration and support; offering customized fine-tuning and optimization services for edge applications.
Competitive Landscape
- OpenAI GPT models
- Google PaLM
- Meta LLaMA
- Cohere
- Anthropic Claude
Implementation Challenges
- Competition from large established AI model providers
- Hardware variability across edge devices complicating optimization
- Balancing model size accuracy and latency for diverse applications
- Adoption inertia in enterprises for new AI architectures
Validation Strategy
- Benchmark LFM2 models against leading edge AI models on latency and accuracy
- Pilot deployments with mobile and IoT device manufacturers
- Collaborate with enterprise customers for real-world use case validation
- Expand multimodal and retrieval capabilities through partner integrations
Research Paper Overview
LFM2 Technical Report
Summary
LFM2 is a family of Liquid Foundation Models optimized for efficient on-device deployment with strong task performance. It features a compact hybrid backbone combining gated short convolutions and grouped query attention blocks, achieving up to 2x faster CPU inference than similar models. The family spans 350M to 8.3B parameters, including dense and mixture-of-experts variants, all supporting 32K context length. Training includes a novel Top-K knowledge distillation, curriculum learning, and a three-stage post-training process. LFM2 models excel on benchmarks like IFEval and GSM8K and extend to multimodal, audio, and retrieval tasks with open weights and deployment tools for edge applications.