Liquid AI
Liquid AI
One-line summary: MIT-spinout foundation-model lab (ramin-hasani, CEO) building Liquid Neural Networks — a continuous-time, post-transformer architecture — into efficient small language models that run on-device (Mercedes-Benz, Shopify, AMD AI PCs).
What it is
Founded ~2023 out of MIT CSAIL (Daniela Rus's lab) by ramin-hasani and Matthias Lechner. Its Liquid Neural Networks (LNN) derive from the continuous-time analog dynamics of the C. elegans worm's 302-neuron nervous system — recurrent, not attention-based. Rather than bet on one architecture, Liquid built an automated architecture-search stack (STAR — "Automated Design of Tailored Architectures," and AFMD — Automated Foundation Model Design) that optimizes per-hardware for memory, compute, latency, and accuracy. Per dave-blundin (early investor) in 2026-07-17-podcast-moonshots-mira-murati-s-975b-open-model-ramin-hasani-on, it is "the only foundation model company that... reached unicorn status coming out of MIT."
Why it matters to artificial-intelligence
Liquid is the thread's clearest instance of the post-transformer-architectures question — and its honest complication: when STAR searches "without a human bias," it re-discovers a hybrid where "double gated convolution" is ~80% of the network with "a little bit of transformers." So Liquid is both the post-transformer bet and evidence that the transformer keeps re-emerging. Its real edge is efficiency at the edge-inference-shift: a <1GB multimodal model on a ~$60 automotive chip, offline and private.
Why it matters to stock-market
Liquid routes AI demand to edge/on-device silicon and names deployment channels: Mercedes-Benz (automotive infotainment/in-car intelligence, rolling out to 2022+ NA cars), Shopify (in production ~6 months, ~10B product requests), Amazon + AMD (AI PCs), plus DOD and a joint regulatory submission with AMD. The tradeable read-through is the on-device chip beneficiaries (Qualcomm/Samsung/AMD) and the OEM adoption pattern, not a listed Liquid ticker (private).
Key facts
- Architecture: Liquid Neural Networks (continuous-time, recurrent); STAR/AFMD automated architecture search.
- On-device: <1GB multimodal models; runs "below the operating system," offline, ~600MB OTA updates.
- Named customers: Mercedes-Benz, Shopify, Amazon/AMD (AI PCs), DOD.
- Private unicorn; MIT spinout (2023).
Open questions
- Is Liquid still meaningfully "post-transformer," or a customized transformer-derivative business? (alexander-wissner-gross's challenge.)
- Does fine-tuning-as-a-service for on-device models scale as a business model?
Sources
- 2026-07-17-podcast-moonshots-mira-murati-s-975b-open-model-ramin-hasani-on (multi-context, vault/sources/)