Liquid AI Tunes On-Device Context Engine for Qualcomm Snapdragon Chips

Liquid AI has adapted its personalization software, Liquid Context, to run natively on Qualcomm Snapdragon processors, the company announced at Qualcomm's Snapdragon Summit in Maui. The software builds and maintains a localized model of user preferences and routines directly on consumer devices, as reported by The Next Web (https://thenextweb.com/news/liquid-ai-liquid-context-qualcomm-snapdragon-vision-dspark).

Spun out of the Massachusetts Institute of Technology in 2023, Liquid AI develops artificial intelligence models intended to run directly on edge hardware—such as smartphones, laptops, connected vehicles, and robots—rather than relying entirely on remote data centers. With user permission, Liquid Context aggregates signals from the device to track schedules and habits, passing relevant pieces of context only to authorized AI agents running either locally or in the cloud.

The context layer runs in the background on Qualcomm's Hexagon Neural Processing Unit (NPU), keeping personal context data stored locally on the device rather than routing routine updates through cloud models. "Personal AI starts with understanding how you live and what you need, when you need it," said Ramin Hasani, chief executive and co-founder of Liquid AI.

The startup outlined illustrative applications for the system, such as an assistant checking calendars and drafting reschedule emails if a child falls ill, or an automotive agent cooling a vehicle cabin based on post-exercise metrics from a connected smartwatch. Qualcomm supplies silicon across the automotive sector and secured an agreement in July to handle vehicle computing for BMW.

Device manufacturers can integrate Liquid Context as a built-in feature and can license Liquid Agent, an operational assistant powered by Liquid AI's 2.6-billion-parameter LFM2.5-2.6B model, which has also been tuned for the Hexagon NPU. While Qualcomm and Liquid AI noted they are exploring further opportunities to collaborate, neither firm named specific device manufacturers currently shipping the software.

"By combining our industry-leading Snapdragon platforms with innovative AI technologies from companies like Liquid AI, we’re helping accelerate the next generation of Agentic AI experiences," said Durga Malladi, executive vice president at Qualcomm Technologies. The announcement comes as Qualcomm broadens its silicon footprint, including developing custom chips for Amazon Web Services, while ecosystem players such as OpenAI work on dedicated devices.

A day after the Snapdragon announcement, Liquid AI released a draft model for LFM2.5-VL-3B, a vision-language model capable of processing both images and text. The system implements speculative decoding, an architecture where a smaller 280-million-parameter draft model proposes upcoming words and the primary model verifies them in a single computational pass to reduce generation latency without altering output quality.

The draft model adds 8.9% to overall model size and was trained exclusively on AMD hardware. According to vendor benchmarks provided by Liquid AI, the approach accelerates decoding by up to 3.13 times on an Apple M5 Max MacBook Pro and up to 2.66 times on an Nvidia H100 GPU, translating to complete response speedups of up to 2.62 times and 2.27 times, respectively.

Liquid AI noted that end-to-end acceleration will be lower across many routine edge tasks because speculative decoding accelerates only text generation. The technique does not reduce the time required to ingest images or process initial prompts, which represent a larger portion of total compute time on resource-constrained devices. The draft model is available for download on Hugging Face.