Liquid AI is a foundation model company building efficient, general-purpose AI at every scale. Founded out of MIT’s CSAIL, we started from a simple observation: 99% of the world’s processors live outside data centers, and come with hard limits on latency, memory, power, and connectivity.
To address this, Liquid designs models around the hardware constraint first, rather than building a model and compressing it to fit afterward. Our Liquid Foundation Models (LFMs) are based on a hybrid architecture that moves past the limitations of transformers, using a hardware-in-the-loop approach with our proprietary Automated Foundation Model Design (AFMD) process to evolve architectures optimized for the exact silicon they’ll run on.
Our view on edge AI’s next unlock: an edge agent isn’t just a smaller cloud agent running locally — it’s a co-developed system, where an efficient, agentic small model, a specialized harness, adaptable memory, and a router work together. Local inference is effectively free between interactions, so an on-device agent can spend that abundant compute reasoning and calling tools more times within the same budget, compounding into a more reliable agent in a way a cloud-metered assistant can’t match. Our recent LFM2.5-2.6B model is built for exactly this — the first edge model capable of reliably completing agentic tasks, fast enough and capable enough to make that thesis real.
Our release from August 2026, LFM2.5-8B-A1B, is a mixture-of-experts model with 8B total parameters but only 1B active at inference time, built for fast local tool calling and agents that can run directly on device. With multimodal models ranging from roughly 230 million to 40 billion-plus parameters, our LFMs already power real deployments across automotive, consumer devices, drug discovery, and enterprise workflows, with partners including Mercedes-Benz, Shopify, and Insilico Medicine.
Backed by an AMD-led $250M Series A at a valuation over $2 billion, Liquid is building the software layer for AI that works in the physical world, not just the data center.
Post by: Hilda Tavakoli, Liquid AI
