Mercedes-Benz taps Liquid AI for on-device voice processing

Under the new partnership, Mercedes says it will bring speech processing, language understanding and reasoning onboard to make interactions faster, more private and more consistent. Liquid’s co-founder and chief executive, Ramin Hasani, an AI and machine-learning scientist from MIT’s Computer Science and Artificial Intelligence Laboratory, says its models can run on hardware already inside the vehicle, according to the companies’ announcement.

Those benefits remain company claims. The announcements provide no finished-system measurements for response time, recognition accuracy, memory requirements or power consumption, and no detailed account of which information will still go to cloud services.

The intended coverage includes combustion-engine and battery-electric vehicles. Mercedes’ eligibility list starts with model-year 2024 onward E-Class and CLE vehicles under third-generation MBUX and model-year 2026 onward electric CLA vehicles under the fourth generation, alongside specified model years of other ranges. That is a coverage plan, not confirmation that listed vehicles have received the software or that existing owners share one update schedule.

Ramin Hasani, center, in the partnership announcement photograph. Image supplied by Liquid AI. Source: Liquid AI.

Integration is the advantage Hasani is selling

Liquid emerged from stealth in December 2023. Its founding announcement identified four MIT laboratory scientists as founders and disclosed $46.6 million in seed funding. In its subsequent Series A announcement, it said it had released text models, announced multimodal models and begun testing products with commercial partners.

Hasani’s pitch goes beyond the model itself. In a September interview published on October 20, 2026, he said working directly with original equipment manufacturers gives Liquid access to vehicle functions. He contrasted that with asking Apple’s Siri to open a car window without the necessary integration. This is an argument about access and supporting software, not proof that a rival assistant could never perform the same task with equivalent integration.

The same report describes Liquid’s work with Shopify, where requests through the Shop app can use a Liquid model. That supplies relevant commercial context beyond automotive, but does not establish the performance of the Mercedes implementation.

Hardware imposes another limit. In a report from Qualcomm’s Snapdragon Summit, Hasani described the Mercedes implementation as running locally on Qualcomm Snapdragon chips. He said the implementation distributes work across several parts of the processor. Embedding the assistant requires fitting the software to the available computing resources, not simply installing a smaller language model.

Hasani called the work a production-grade deployment at the summit. In the September interview, however, he still expected a North American rollout by the end of 2026. Neither account establishes how many customers had received it. His own summit remarks also identified a gap between device models and frontier cloud models as an obstacle to wider adoption of local AI.

Liquid AI raised $250 million in an AMD-led round in December 2024. The figure represents financing, not the value of its Mercedes-Benz agreement. Graphic: Magica. Source: Reuters via Investing.com.

Financing efficiency still requires computing infrastructure

Liquid said in December 2024 that it had closed a $250 million funding round led by chipmaker Advanced Micro Devices, also its strategic partner, according to a report on the financing. The backing connects Liquid’s efficiency strategy to a chip supplier with a stake in its deployment.

Liquid’s Series A announcement said the money would expand computing infrastructure and accelerate readiness for inference and fine-tuning on devices and customers’ premises. AMD said it would collaborate on training and deploying Liquid models on AMD Instinct GPUs. Local inference—the execution of a model—therefore does not eliminate the computing infrastructure needed to develop and adapt it.

Nor does it remove licensing costs. The October interview report describes Liquid’s general business model as annual fees per device plus development costs for manufacturers, and token-based charges for enterprise model customization. Hasani said potential annual licensing revenue could range from less than $1 to hundreds of dollars per device, depending on the application. Those figures are not disclosed Mercedes prices, and the report says Liquid has not disclosed revenue forecasts.

Moving tasks onboard can reduce dependence on remote-server requests, but the Mercedes announcements provide no contract value, per-vehicle fee or measured total-cost comparison. The available evidence cannot establish net savings after licensing, integration and local computing costs.

The next test is availability and reliable vehicle control

The immediate evidence needed is a vehicle-by-vehicle release schedule and confirmation of the functions customers can use without connectivity. The published production target and Hasani’s year-end expectation set a window; they do not demonstrate delivery.

The more consequential test is whether the local system reliably completes vehicle commands within the hardware available. Response times, command-completion accuracy, offline coverage and a clear explanation of data sent to the cloud would let customers assess the promised benefits. A disclosed cost comparison would let manufacturers assess the economics. Until then, the partnership establishes a route to embedded AI, while its advantage in everyday use remains to be demonstrated.