Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

2026-07-04 · Source: The Cognitive Revolution · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Internet of Things (IoT) & Connected Devices · Depth: Advanced, short

Summary

Liquid AI, founded by MIT researchers, is developing device-native foundation models optimized for edge devices with limited memory and processing power. Their initial research involved biologically inspired neural networks, demonstrating capabilities like parking a car with just 12 liquid neurons. Today, Liquid AI maintains a neutral, empirical approach, holding the number five spot on the Hugging Face United States total downloads leaderboard and partnering with companies like Shopify and Mercedes-Benz. Their Apollo app showcases a one-billion-parameter model running efficiently on an iPhone for basic use cases such as local document search. The company's network architecture search process evaluates models on real downstream tasks and target hardware, often leading to exotic architectures like Mamba for specific, resource-constrained applications, while attention-based models still dominate the frontier. Liquid AI plans to introduce a self-serve platform for fine-tuning small models.

Key takeaway

For ML engineers developing edge AI solutions, you should consider Liquid AI's empirical approach to network architecture search, which prioritizes real-world performance on target hardware. Explore their Apollo app or upcoming fine-tuning platform to optimize models for resource-constrained environments. This strategy can reduce reliance on large frontier models, improve data privacy, and expand AI accessibility globally, especially for specific use cases where exotic architectures excel.

Key insights

Optimizing neural networks for edge devices requires empirical evaluation on target hardware and specific use cases.

Principles

• Evaluate models on real downstream tasks and target hardware.
• Specific use cases and limited compute favor exotic architectures.
• Attention-based architectures generalize best for frontier models.

Method

Liquid AI employs a network architecture search process that directly evaluates models on actual target hardware for specific downstream tasks, moving beyond proxy metrics.

In practice

• Run 1B-parameter models on iPhones for local document processing.
• Explore Mamba-like architectures for resource-constrained edge deployments.
• Utilize self-serve platforms for fine-tuning small models.

Topics

• Edge AI
• Device-Native AI
• Foundation Models
• Neural Network Architectures
• Model Optimization
• Liquid AI
• Mamba Architecture

Best for: Investor, AI Architect, AI Engineer, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• The Sequence Knowledge #870: Liquid Models and the Search for a Post-Transformer Architecture — Transformers excel at cloud-scale AI but are costly for on-device applications, where liquid models offer a dynamic, efficient alternative.
• The new enterprise AI expert every company needs - and why — Frontier engineers with deep data and neural networking expertise are vital for optimizing enterprise AI models.
• Apple's Edge Moat & the AI Frontier Gap — Apple's AI strategy requires both competitive edge inference and strong frontier models to succeed against cloud-based offerings.
• Neural Engines Finally Explained (In 5 Minutes) — Domain-specific hardware like Neural Engines optimizes AI inference by prioritizing spatial architecture and data quantization.
• AI Weekly Issue #508: The Cutting Edge, Across the Board — AI advancements span from massive open models to efficient edge deployments, increasingly impacting real-world applications and the economy.
• Cognichip wants AI to design the chips that power AI, and just raised $60M to try — AI can significantly accelerate and de-risk complex chip design by automating and optimizing engineering workflows.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Cognitive Revolution.