Liquid AI 已对其软件 Liquid Context 进行了优化，使其能够在高通（Qualcomm）的 Snapdragon 芯片上运行。该软件能够在设备本身上收集并分析用户的各种信息（如日常习惯、偏好等），并持续更新这些信息。该公司于周三在毛伊岛举行的高通 Snapdragon Summit 上宣布了这一进展。
Liquid AI has tuned Liquid Context to run on Qualcomm’s Snapdragon chips. The software builds a picture of the user on the device itself. The company announced the work at Qualcomm’s Snapdragon Summit in Maui on Wednesday. A day later, it released a tool that speeds up its vision model on devices. Liquid AI, spun out of MIT in 2023, builds AI models meant to run on phones, laptops, cars and robots rather than in data centres.
一天后，Liquid AI 发布了一款工具，用于进一步提升其视觉处理模型的运行效率。该公司成立于 2023 年，专注于开发适用于手机、笔记本电脑、汽车及机器人等设备的 AI 模型，而非仅限于数据中心环境。
What Liquid Context does With the user’s permission, Liquid Context learns routines, preferences and needs from signals on the device. It keeps that picture up to date. It then passes the relevant parts to the AI agents the user chooses. Those can be Liquid AI’s own Liquid Agent or other companies’ agents. They can run on the device, in the cloud or both.
在用户的许可下，Liquid Context 会从设备上收集相关信息并学习用户的习惯与需求，然后将其传递给用户选择的 AI 代理（这些代理可以是 Liquid AI 自己开发的，也可以是其他公司的产品）。这些 AI 代理可以在设备上运行，也可以在云端运行，或者同时运行在设备和云端。
What Liquid Context does With the user’s permission, Liquid Context learns routines, preferences and needs from signals on the device. It keeps that picture up to date. It then passes the relevant parts to the AI agents the user chooses. Those can be Liquid AI’s own Liquid Agent or other companies’ agents. They can run on the device, in the cloud or both.
该公司表示，用户的个人信息会始终存储在设备上，AI 代理仅能访问用户允许访问的数据。该软件运行在 Qualcomm 的 Hexagon NPU（专为 AI 设计的芯片模块）上，因此无需依赖云端服务器来处理所有数据更新。
The company said the personal context stays on the device. Agents only get what the user allows. The layer runs in the background on the Hexagon NPU, Qualcomm’s chip block for AI. That way, a cloud model does not have to process every update.
“个人化的 AI 服务首先需要了解用户的生活方式及其需求，并在用户需要时提供相应的帮助，”Liquid AI 的首席执行官兼联合创始人 Ramin Hasani 如是说。Liquid AI 提供了三个示例来说明其技术的应用场景：
“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 gave three examples, which it called illustrative. In one, a message says the user’s child is sick and needs collecting from school early. An agent checks the calendar, works out which meetings can move and drafts emails for the user to approve.
当系统检测到用户的孩子生病、需要提前从学校接孩子时，相关代理会查看日程安排，调整会议时间，并为用户生成需要审批的电子邮件；
Liquid AI gave three examples, which it called illustrative. In one, a message says the user’s child is sick and needs collecting from school early. An agent checks the calendar, works out which meetings can move and drafts emails for the user to approve.
当用户跑步时，智能手表会记录下他们的个人最佳成绩，用户回到车内后，车载代理会向他们表示祝贺，并根据用户的喜好调节车内温度；
In another, a smartwatch records a personal best on a run. When the user gets back to the car, the car’s agent congratulates them and cools the cabin to their liking. Qualcomm already supplies chips to carmakers, and BMW handed it its cars’ computing in July.
高通已向汽车制造商提供相关芯片，宝马公司已于 7 月份在其汽车系统中采用了 Liquid AI 的技术。
In another, a smartwatch records a personal best on a run. When the user gets back to the car, the car’s agent congratulates them and cools the cabin to their liking. Qualcomm already supplies chips to carmakers, and BMW handed it its cars’ computing in July.
总之，Liquid AI 的目标是通过智能软件帮助用户更便捷地满足日常生活中的各种需求。
What Qualcomm gets Device makers can build Liquid Context in as a standard feature, and can also license Liquid Agent. That agent runs on LFM2.5-2.6B, a model with 2.6 billion parameters. Liquid AI has tuned it for the Hexagon NPU. The two companies said they are “exploring additional opportunities” to work together. Neither company named a device maker.
高通获得的收益 设备制造商可将 Liquid Context 作为标准功能内置，也可授权使用 Liquid Agent。该 Agent 运行在 LFM2.5-2.6B 上，这是一个拥有 26 亿参数的模型。Liquid AI 已针对 Hexagon NPU 对其进行了优化。双方表示正在“探索更多合作机会”。两家公司均未透露设备制造商的名称。
What Qualcomm gets Device makers can build Liquid Context in as a standard feature, and can also license Liquid Agent. That agent runs on LFM2.5-2.6B, a model with 2.6 billion parameters. Liquid AI has tuned it for the Hexagon NPU. The two companies said they are “exploring additional opportunities” to work together. Neither company named a device maker.
“通过将我们行业领先的骁龙平台与 Liquid AI 等公司的创新 AI 技术相结合，我们正在帮助加速下一代智能体 AI 体验的到来，”高通技术公司执行副总裁 Durga Malladi 表示。
“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.
高通还在为数据中心制造芯片，包括为亚马逊 AWS 定制的芯片。与此同时，OpenAI 正在研发自己的设备。
Qualcomm is also making chips for data centres, including custom silicon for Amazon’s AWS. OpenAI, meanwhile, is working on its own devices.
更快的视觉模型 周四，Liquid AI 发布了 LFM2.5-VL-3B 的草稿模型。该模型既能读取图像也能处理文本。它采用了一种称为推测解码的方法。一个小型草稿模型预测接下来的几个词，主模型一次性验证它们，从而在保持同等质量的前提下加快输出速度。
A faster vision model On Thursday, Liquid AI released a draft model for LFM2.5-VL-3B. That model reads images as well as text. It uses a method called speculative decoding. A small draft model guesses the next few words, and the main model checks them in one pass, so the answer comes out faster with the same quality.
草稿模型拥有约 2.8 亿参数，使模型体积增加了 8.9%。Liquid AI 称，在 M5 Max MacBook Pro 上解码速度最高可提升 3.13 倍，在 Nvidia H100 GPU 上最高可提升 2.66 倍。完整回答的生成速度分别最高可提升 2.62 倍和 2.27 倍。该公司表示仅使用 AMD 硬件进行了训练。
The drafter has about 280 million parameters and adds 8.9% to the model’s size. Liquid AI said it speeds up decoding by up to 3.13 times on an M5 Max MacBook Pro and up to 2.66 times on an Nvidia H100 GPU. Whole answers arrive up to 2.62 and 2.27 times faster. The company trained it only on AMD hardware.
Liquid AI 表示，对于许多任务，实际提升幅度小于表面数字。该方法无法加速图像或提示词的读取过程。在边缘设备上，这些步骤占总耗时的比例更大，从而限制了整体加速效果。该模型可在 Hugging Face 上免费下载。
The gains are smaller than they look for many tasks, Liquid AI said. The method does not speed up reading the image or the prompt. On edge devices, those steps take a bigger share of the total time, which caps the overall speed-up. The model is free to download on Hugging Face.