Liquid AI Integrates Personal Context Layer for AI Agents on Snapdragon Chips
Liquid AI has tuned Liquid Context, which learns about the user on the device, for Qualcomm's Snapdragon chips. A day later, they released a faster vision model.
September 24, 2026 - 5:58 pm Credit: Liquid AI
Liquid AI has optimized Liquid Context to run on Qualcomm’s Snapdragon chips. The software creates a user profile directly on the device. The company announced this at Qualcomm’s Snapdragon Summit in Maui on Wednesday.
A day later, they introduced a tool that accelerates their vision model on devices. Liquid AI, spun out of MIT in 2023, develops AI models designed to run on phones, laptops, cars, and robots instead of in data centers.
What Liquid Context Does
With the user’s permission, Liquid Context learns routines, preferences, and needs from signals on the device. It keeps this information updated and shares relevant details with AI agents chosen by the user. These agents can be Liquid AI's own Liquid Agent or those from other companies, operating on the device, in the cloud, or both.
The personal context remains on the device, and agents only access what the user allows. It runs in the background on Qualcomm’s Hexagon NPU, an AI chip block. This way, a cloud model doesn't need to process every update.
“Personal AI starts with understanding how you live and what you need, when you need it,” said Ramin Hasani, Liquid AI’s chief executive and co-founder.
Liquid AI provided three examples:
-
A message informs the user that their child is sick and needs to be collected from school early. An agent checks the calendar, reschedules meetings, and drafts emails for user approval.
-
After a smartwatch records a personal best during a run, the car’s agent greets the user upon their return, congratulates them, and adjusts the cabin temperature to their preference.
Qualcomm already supplies chips to various carmakers, and BMW integrated it into its cars' computing in July.
What Qualcomm Gains
Device manufacturers can integrate Liquid Context as a standard feature and also license Liquid Agent. This agent runs on LFM2.5-2.6B, a model with 2.6 billion parameters optimized for the Hexagon NPU. The companies stated they are exploring further collaborations.
“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, an executive vice president at Qualcomm Technologies.
Qualcomm also produces chips for data centers, including custom silicon for Amazon’s AWS. OpenAI is developing its own devices, too.
A Faster Vision Model
On Thursday, Liquid AI released a draft model, LFM2.5-VL-3B, capable of reading both images and text. It employs a method called speculative decoding: a small draft model guesses the next few words, which the main model checks in one pass for faster, high-quality output.
The drafter has approximately 280 million parameters, increasing the model's size by 8.9%. Liquid AI reported significant speedups on an M5 Max MacBook Pro (up to 3.13 times) and Nvidia H100 GPU (up to 2.66 times). Whole answers arrive up to 2.62 and 2.27 times faster. Trained only on AMD hardware, the gains are less notable for certain tasks, as this method doesn't accelerate reading the image or prompt.