Liquid AI

What Liquid AI does

Liquid AI builds efficient, general-purpose AI at every scale, focusing on device-native foundation models (LFMs) designed for the latency, privacy, and hardware constraints of the physical world. Their mission is to provide highly capable, compute and cost-optimized AI that can run on any device and medium. They offer a full-stack solution including architecture, optimization, and deployment engines to accelerate the path from prototype to product, enabling rapid customization and deployment of models on edge devices. Liquid AI also provides the Liquid Edge AI Platform (LEAP) SDK for developers to specialize, deploy, and fine-tune models.

Products

• Liquid Foundation Models (LFMs)Device-native foundation models built for latency, privacy, and hardware constraints. They are designed for rapid customization and peak performance for specified use cases at a small footprint, running locally on chosen hardware.
• LEAP SDK (Liquid Edge AI Platform SDK)A software development kit that provides the fastest path from notebook to production, allowing users to fine-tune models, bake them to any runtime, and ship them. It includes tools for model search, testing, comparison, customization (fine-tuning CLI), model bundling, and an Edge SDK for integration.
• Liquid ApolloA platform to download and test LEAP models on your own device, or use a cloud playground for testing and comparison.
• LFM2-1.2BA text-based Liquid Foundation Model with 1.2 billion parameters.
• LFM2-Audio-1.5BAn audio-based Liquid Foundation Model with 1.5 billion parameters.
• LFM2-350MA Liquid Foundation Model with 350 million parameters.
• LFM2-VL-1.6BAn image-based Vision-Language Liquid Foundation Model with 1.6 billion parameters.
• LFM2-700MA Liquid Foundation Model with 700 million parameters.
• LFM2-VL-450MA Vision-Language Liquid Foundation Model with 450 million parameters.
• LFM2-350M-ENJP-MTA Liquid Foundation Model with 350 million parameters, specialized for English-Japanese Machine Translation.
• LFM2-1.2B-RAGA Liquid Foundation Model with 1.2 billion parameters, specialized for Retrieval Augmented Generation.
• LFM2.5-230MA Liquid Foundation Model built to run anywhere.
• LFM2.5 RetrieversBi-directional LFMs for fast multilingual search.
• LFM2.5-8B-A1BAn on-device Mixture of Experts Liquid Foundation Model.
• LFM2.5-VL-450MA Vision-Language Liquid Foundation Model for structured visual intelligence, from edge to cloud.
• LFM2.5-350MA Liquid Foundation Model with 350 million parameters.
• LFM2-24B-A2BA large LFM2 model for tool-calling agents on consumer hardware and scaling up the LFM2 architecture.
• LFM2.5-1.2B-ThinkingAn on-device reasoning model under 1GB.

Key capabilities

• Device-native foundation models
• Advanced intelligence for processors outside of data centers
• Built for latency, privacy, and hardware constraints
• 40 LFMs shipped, 40.7M downloads, 2800+ variants
• Efficient, general-purpose AI at every scale
• Highly capable, compute and cost optimized
• Rapid customization for peak performance
• Full-stack solution: architecture, optimization, deployment engines
• Hardware-optimized VLMs
• Reduced model size without sacrificing accuracy
• Real-time AI interactions directly on devices
• Tailored, multimodal AI on any device, at industrial scale
• Works fully offline on all devices
• Real-time, hands-free assistance for workers
• Agentic workflows for IT/OT orchestration
• Reasons across SCADA, PLCs, and enterprise systems
• Real-time analysis, predictive maintenance, autonomous workflows
• Memory-efficient models running on edge CPUs, NPUs, GPUs
• Sub-millisecond reflexes and 24/7 reliability for robotics
• Low latency and bank-grade privacy for financial services
• Hosted entirely on your hardware for inherent privacy
• Tailored small models for specific functions
• Lightning-fast response times for digital and mobile customers
• Compute-optimized
• Run anywhere
• Specialized
• Privacy-forward
• Best model search
• Model library
• Vision-enabled models
• Fine-tuning tools
• Model bundling service
• Edge SDK for seamless integration
• Laptop support
• Function calling
• Efficient training via packing
• Custom embedding & inference stack optimized for CPU
• Compact state representation
• Privacy-focused deployment (on-premise/on-tenant)

Use cases

• Phones, laptops, cars, space, e-commerce, financial services, bio, defense deployments
• Real-time voice and vision AI in vehicles
• Industrial automation and robotics
• Empowering front-line workers with customized, human-level reasoning
• IoT sensors, mobile computers, AR headsets made more useful
• Agentic IT/OT orchestration for insights and efficiencies
• Predictive maintenance
• Autonomous workflows
• Robotic physical autonomy
• Real-time financial intelligence from transactions and signals
• Fraud detection
• Customer service in finance
• Risk functions in finance
• On-Device Scene Understanding in AR/VR
• On-Device Low-Latency Dialogue for Games
• On-Device Real-Time Fraud Detection
• Offline Knowledge Assistant
• On-Prem Private AI for Enterprises
• Edge Predictive Maintenance
• On-Device Health Analytics for Wearables
• On-Prem Document Processing
• In-Vehicle Function Orchestration
• On-Prem Compliance Automation
• Product cataloging for retail
• Real-time, on-device translation for smartphones
• Drug discovery (scientific foundation models)

AI approach

Liquid AI builds device-native foundation models (LFMs) designed for efficient deployment on edge devices with latency, privacy, and hardware constraints. They offer a platform (LEAP SDK) for fine-tuning, optimizing, and deploying these models across various runtimes and hardware. Their approach focuses on creating specialized, small models that perform exceptionally well for specific use cases, often outperforming larger generic models, while being compute and cost-optimized.

Tech named: Liquid Foundation Models (LFMs), Liquid Edge AI Platform (LEAP), LEAP SDK, llama.cpp, MLX, ONNX, CoreML, SGLang, vLLM, Liquid neural networks, state-space models, Liquid Apollo, GRU model

Industries served

• Automotive
• Industrial & Robotics
• Financial Services
• E-commerce
• Bio (Biotechnology)
• Defense
• Retail
• Consumer Electronics (Smartphones)
• Healthcare (Wearables)
• Space

What it says sets it apart

• Device-native foundation models optimized for edge constraints
• Efficiency-first approach: compute and cost optimized
• Rapid customization and deployment (LEAP SDK)
• Full-stack solution from architecture to deployment
• Hardware-optimized models for existing CPUs
• Significant performance improvements (e.g., 10x faster TTFT, 50% smaller models)
• Offline capabilities for security and latency
• Agentic workflows for complex industrial and financial tasks
• Inherent privacy by design (on-device/on-premise deployment)
• Tailored small models for specific functions, outperforming larger generic models
• Spun out of MIT CSAIL with founders from MIT
• Focus on state-of-the-art, general-purpose AI systems that are capable, efficient, highly aligned, and trustworthy

Funding rounds we track

From the AI funding tracker — rounds as reported by the linked publications.