Learn About Liquid AI's AI Products
Create a free account to access in-depth lessons on each tool and model.
Start Learning Free📋About Liquid AI
Updated June 15, 2026Liquid AI is a Cambridge, Massachusetts foundation-model lab spun out of MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) in 2023. The four founders — Ramin Hasani (CEO), Mathias Lechner, Alexander Amini, and Daniela Rus (the MIT roboticist who directs CSAIL) — built the company around their academic work on liquid neural networks, a class of continuous-time models originally developed for robotic control and adapted at Liquid into the Liquid Foundation Model (LFM) line.
The company's thesis is that the next decade of generative AI will run primarily on-device — laptops, phones, automotive ECUs, industrial controllers — rather than in hyperscaler data centers, and that the dominant transformer architecture is poorly matched to that constraint. Liquid's LFM line is engineered around two design choices that follow from that thesis: a sparse mixture-of-experts (MoE) architecture that keeps active parameters small relative to total parameters, and an open-weight release strategy that lets customers fine-tune and deploy without per-token API fees.
The headline 2026 release is LFM2.5-8B-A1B — 8 billion total parameters, roughly 1 billion active per token, pretrained on 38 trillion tokens, with a 128,000-token context window and an open-weight license that imposes no use restrictions. Liquid reports inference at roughly 253 tokens per second on an Apple M5 Max, 146 tokens per second on a Ryzen AI Max Plus, and around 30 tokens per second on flagship smartphones — putting credible local inference in the same parameter band as the closed flagship-mini tier. The model ships day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX, removing the integration friction that has historically slowed open-weight adoption.
Beyond the LFM line itself, Liquid sells the LEAP customization and deployment platform for enterprise teams who want to fine-tune LFM checkpoints against private data and ship the results to their own devices, plus Liquid Apollo, a consumer-facing on-device AI application. The combined pitch reframes Liquid against frontier-lab competitors: rather than competing on capability ceiling (where Anthropic, OpenAI, and Google maintain large leads), Liquid competes on cost-per-token at the edge, on data residency, and on the simple fact that an on-device model continues to work when the network does not.
The broader bet matters because hyperscaler-hosted inference has emerged as the single largest cost line for AI-product companies, and on-device inference at flagship-mini quality changes that calculation for any application that does not need the absolute frontier — keyboard agents, in-car voice, manufacturing-floor copilots, accessibility assistants, embedded coding helpers. Liquid is one of a small number of labs (alongside Apple's on-device Foundation Models, Microsoft's Phi family, and Google's Gemma line) building explicitly for that constraint.
🛠️Products & Tools (1)
Liquid Foundation Models (LFM) — open-weight mixture-of-experts models engineered for on-device inference. Current flagship LFM2.5-8B-A1B has 8 billion total parameters with roughly 1 billion active per token, was pretrained on 38 trillion tokens, and runs with a 128,000-token context window on a laptop or a flagship smartphone via llama.cpp, MLX, vLLM, SGLang, or ONNX. Open-weight with no use restrictions; ideal when you need flagship-mini quality at the edge instead of in a hyperscaler data center.