MAUI, Hawaii — Liquid AI announced a new on-device memory system at Snapdragon Summit 2026. The system promises to change how AI assistants interact with users. No longer limited to one-off queries, these assistants could soon anticipate needs based on routines built from phone sensors, calendars, and habits.

Liquid Context Arrives on Snapdragon Hardware

The partnership pairs Liquid AI’s context layer with Qualcomm’s Hexagon NPU. Liquid Context runs continuously in the background. It gathers signals with user permission. Then it compresses those signals into a persistent understanding of preferences and daily patterns. That understanding stays local. Agents — whether running on the device, in the cloud, or in a hybrid setup — tap into it as needed.

Ramin Hasani, CEO and co-founder of Liquid AI, put it plainly. “Personal AI starts with understanding how you live and what you need, when you need it.” He continued, “Liquid Context builds that understanding on your device so the agents you choose can offer more relevant help and anticipate your needs. The Hexagon NPU makes this continuous, local intelligence practical. Together, we are bringing personal context to the devices people rely on every day.” (Liquid AI)

Current assistants forget quickly. They react. They rarely predict. A user might ask for a smoothie recipe one day. The next morning the same assistant suggests one based on recent exercise data, preferred ingredients tracked from past grocery lists, and the time of day. But the important part isn’t the smoothie. It’s that the suggestion follows the user from phone to laptop to car without repeated cloud uploads of sensitive details.

Qualcomm executives see the same gap. Durga Malladi, executive vice president at Qualcomm Technologies, said the collaboration accelerates “the next generation of Agentic AI experiences.” Both companies plan to deepen ties so OEMs can offer personal context as a built-in feature. (Liquid AI)

But. This isn’t just another memory cache. Liquid Context acts as a shared layer. Third-party agents get relevant slices of context. Liquid Agent, powered by the optimized LFM2.5-2.6B model, also benefits. The 2.6-billion-parameter agentic model runs efficiently on the Hexagon NPU thanks to targeted changes in memory access and latency. Scalar, vector, and matrix processing combine with dedicated transformer hardware to handle the workload without draining battery or spiking heat.

Privacy questions remain. Data stays on device for processing. Yet permitted agents can still receive context snippets. Users control permission. The architecture avoids constant cloud round-trips for every background update. That matters for battery life and data costs. It also addresses growing regulatory pressure around personal information.

The timing fits Qualcomm’s broader message. Cristiano Amon, Qualcomm CEO, has called 2026 the year of agents. He argues agents will become the new computing platform. Five conditions must hold: model quality must close the gap with cloud systems, context must travel with the user, better tools must surround the models, physical limits like battery and heat must be respected, and the system must work across devices. Liquid Context targets the context condition directly. (NERDS.xyz)

Liquid AI didn’t stop at context. One day after the Snapdragon announcement the company released LFM2.5-VL-DSpark, an experimental draft model for its 3B-parameter vision-language model. The 280-million-parameter drafter adds just 8.9% more parameters. In return it delivers decoding speedups up to 3.13× on edge devices and 2.66× on GPUs. End-to-end throughput gains reach 2.62× on edge hardware. The model works with llama.cpp, SGLang, and MLX-VLM from day one. (Liquid AI, published Sept. 24, 2026)

Speculative decoding for vision-language models brings its own limits. Vision tokens differ from text. The drafter taps hidden states from the target model at fixed layers, then proposes blocks of candidate tokens. Acceptance rates determine real gains. On certain edge chips the gains shrink when image processing dominates. Still, the numbers look promising for automotive and wearable use cases where real-time vision matters.

Automakers already test similar technology. Liquid AI previously helped a global carmaker deploy real-time voice and vision AI. The optimized model ran 10 times faster on existing hardware. Model size dropped 50% with no accuracy loss. Deployment time fell from months to one week. Mercedes-Benz appears in early Snapdragon Digital Cockpit work with Liquid AI models. (Liquid AI)

Device makers now hold new options. They can embed Liquid Agent as their own branded assistant. Or they can expose the context layer to any agent the user prefers. The memory layer compresses experiences across phones, laptops, wearables, and cars. Context travels. The agent on the wrist knows what the phone learned earlier that day.

Challenges persist. Model quality still trails the best cloud systems in some tasks. Continuous background processing demands careful thermal and power management even on advanced NPUs. User trust around always-on monitoring will decide adoption rates. Yet the combination of persistent local memory and faster inference tools moves the industry closer to agents that feel truly personal.

Hasani’s MIT roots show in the approach. His team draws inspiration from efficient biological systems. The result is a family of Liquid Foundation Models designed from the start for edge constraints rather than scaled down from data-center giants. LFM2.5 models already see 1.4 million downloads weekly on Hugging Face. The open-weight strategy invites developers to experiment while Liquid AI monetizes through OEM licensing and enterprise deployments.

Analysts note the shift. Personal context could become the moat that decides which devices win loyalty. A phone that remembers your routines and anticipates needs without constant cloud dependence offers sticky value. Qualcomm’s Snapdragon platform reaches hundreds of millions of devices. Liquid AI’s software now has broad distribution potential.

So the announcement marks more than a technical integration. It signals a bet that the next leap in AI usefulness comes less from bigger models and more from models that know their user intimately — and respect their device boundaries. OEMs, developers, and users will decide if the bet pays off. Early signs suggest many will want to try.