Problem

Liquid AI addresses the deployment gap created by conventional generative AI: powerful models typically require substantial compute, a cloud connection, and recurring per-token charges. That is a poor fit for phones, cars, industrial controllers, and privacy-sensitive enterprise workflows where latency, connectivity, data sovereignty, memory, and energy consumption matter. Liquid’s own pricing materials emphasize that local inference can operate in milliseconds, work offline, and avoid recurring usage bills, while its research describes embedded devices that often lack a spare GPU and cannot afford a cloud round trip.

The clearest use case is private, real-time intelligence at the edge: an in-car assistant, local copilot, device-based productivity agent, or enterprise application handling sensitive data on-premise. Mercedes-Benz’s 2026 partnership is explicitly aimed at real-time, private, local AI for onboard services, while Liquid’s models are also being applied to pharmaceutical research, where on-premise execution can keep scientific data within a private infrastructure.

Product / Service

Liquid AI builds device-native Liquid Foundation Models, or LFMs: compact, customizable foundation models designed to deliver strong general-purpose and multimodal performance with much less compute than large cloud models. Its LFM2 family uses a hybrid architecture optimized for local and edge inference, with reported gains in training and CPU decode speed, and the models are designed to run across CPUs, GPUs, and NPUs on smartphones, laptops, vehicles, and other hardware. Newer LFM2.5 releases extend the product line to on-device agents, tool calling, encoders, and mixture-of-experts models.

The delivery model combines open-weight models with a deployment stack and enterprise services. Developers can download, run, and fine-tune open LFMs, including commercially below the company’s $10 million annual-revenue threshold, while larger enterprises can purchase commercial licensing, bespoke optimization, OEM and on-premise deployment support, and service-level agreements. Liquid’s LEAP platform simplifies deployment to iOS and Android, and the company supports common runtimes such as llama.cpp, MLX, vLLM, SGLang, and ONNX. The benefit is a path from model customization to production that reduces cloud dependence while improving latency, privacy, and operating economics.

Market

Liquid AI competes in the small and efficient foundation-model, edge-AI, and on-device inference markets. Its competitive set includes Qwen3 and other small models from Alibaba and ByteDance, as well as adjacent open or device-oriented models such as Apple’s OpenELM and the broader Phi, Gemma, and Llama families. Liquid differentiates through an efficiency-first architecture, optimization for local hardware, open licensing, and a full-stack combination of models, inference runtimes, and deployment tooling rather than competing only as a cloud API provider.

The company has substantial financing and meaningful commercial signals, although the evidence does not disclose revenue. Liquid announced a $250 million Series A in December 2024 led by AMD Ventures, said it was engaged with a large number of Fortune 500 companies, and has announced relationships involving Shopify, Mercedes-Benz, Alef Education, and Insilico Medicine. Its models are openly available through Hugging Face, while the Mercedes-Benz and Insilico partnerships demonstrate production-oriented applications in embedded automotive intelligence and private pharmaceutical research; Liquid therefore appears beyond the purely pre-product stage, but its precise revenue and customer-conversion levels remain undisclosed.