Liquid AI puts personal context at the edge
Key points
Key takeaways from theCUBE interview with Liquid AI COO Jeffrey Li (September 2026):
Liquid Context with Qualcomm. Liquid announced a personal context layer built with Qualcomm:it sits between every model, agent, and hardware, helping the device build a real understanding of the user — who you are with, what you are trying to get done, how you are feeling, and where you have been — information invisible to cloud agents, so on-device agents can act more proactively.
Edge rearchitecture. Cloud-built agent harnesses assume unlimited storageand compute, building ever-growing text-file memory logs pushed to object stores. At the edge those assumptions break: a phone has fixed, zero-sum compute, so Liquid re-divides responsibility for understandingthe user between the harness, model, and a context layer, compressing what it stores into formats readable by the agent, not necessarily by a human.
Non-transformer models. Liquid avoids the exploding KV cachethat keeps transformer-based models off devices. Trade-off: slightly weaker general-purpose capability in exchange for production qualitythat fits device constraints — a PhD in what matters, not a high-schooler in everything — backed bya model portfolio spanning roughlya 100-1000x range in size to target each edge device class per industry.
Edge-vs-cloud routing. Routing heuristics decide where inference runs: low connectivity, data privacy,and ultra-low latency demand on-device computing,and even trivial queries like1+1 should run locally,becausea cloud round-trip costs more than the query's economic value. Keeping data on-device also makes governanceand security fundamentally simpler when designed from first principles.
CoreWeave as training partner. Liquid trains its device-native foundation models on CoreWeave's compute clusters —a bespoke training stack for ultra-specific, alternative-architecture models destined to run outside data centres:the partnership is essentially100% training.
From cars to self-healing agents. Liquid has publicly announced edge agents deployed in Mercedes-Benz cars that take action on the driver's behalf. Next: systems that can self-heal, improveand personalise on their own, with observabilityand continuous-improvement loops baked into the product.
2027 scaling outlook. Li expects 2027 to bea big year: automotive is checked off,and telecom — nervous but waiting near the water with Starlink around the corner— could jump in at MWC. Blockers to full-scale edge deployment: quality — smaller models mean less capability, so Liquid maximises production quality at the edge —and integration, since most industries have no edge foundation model to swap in today: adoption is really a migration from non-genAI systems.
Silicon partnersand momentum. Liquid trains on Nvidia and AMD chips, with AMD one of its largest investors, plus specialist vertical silicon partners. The ~120-person team also just released its first decision model — described asthe first to beat Jev at its own game — available online.
Read more: YouTube · theCUBE event page
