What Liquid AI does
Liquid AI is an efficiency-first foundation model company focused on “liquid”/non-Transformer model directions and related training and inference techniques, with the explicit goal of making general-purpose AI usable across devices and environments (local/edge, on-prem, and hybrid). Liquid positions its models as compute- and cost-optimized while maintaining capabilities for instruction following and agentic workloads, and it emphasizes privacy-forward deployment options (e.g., running without sending confidential data or inference to a third party).
More
Product-wise, Liquid provides (1) Liquid Foundation Models (LFMs) in multiple model families/sizes and modalities, including open-weight checkpoints; (2) LEAP (Liquid Edge AI Platform), described as an SDK/platform for finding, specializing, and deploying models on-device; and (3) Liquid Apollo, described as a local, secure playground for interacting with models on a phone. Liquid’s strategy blends model research and systems/productization: it publishes frequent LFM releases (language, vision-language, audio, retrievers/encoders for retrieval and understanding, and quantization/speculative-decoding variants), and it pairs those releases with tooling and deployment guidance via LEAP and related developer documentation. For example, the company has released quantization-aware distillation checkpoints (e.g., LFM2.5 Q4_0) aimed at keeping edge latency and memory low while recovering accuracy, and it has released speculative-decoding draft checkpoints (DSpark) plus upstream integration references for popular runtimes. Commercially, Liquid highlights partnership-led deployment pathways into production environments where low latency, offline/privacy needs, and hardware constraints matter. Examples include a multi-year collaboration with Mercedes-Benz for embedded in-car intelligence (targeting a production timeline in 2H 2026, per the press release), and a partnership with MacPaw to co-develop on-device AI for the Mac (MacPaw describes the collaboration as aimed at co-developing a tech stack for local AI on Mac). As of late 2026, Liquid continues to expand its model ecosystem and edge deployment story via frequent new releases and supporting tools/benchmarks, including Pipette (an open-source benchmarking suite for edge/on-device intelligence) that Liquid describes as measuring device-class performance across configurations and partnering with Artificial Analysis for methodological validation.