PATEO's AI Box runs NVIDIA Nemotron and ByteDance Doubao on DRIVE AGX Thor

PATEO's announcement, made at Auto China 2026 in Beijing, builds on joint development of the solution with ByteDance and a leading new energy vehicle automaker. The architecture reflects a deliberate split: a compact 2-billion-parameter model handles semantic classification and request rejection, while the 35-billion-parameter Doubao mixture-of-experts model handles the heavier language work, all alongside NVIDIA Nemotron 3.5 ASR for speech recognition.

The design trade-off is memory. As conversation turns and agent actions accumulate during a trip, the model's key-value cache grows, raising demands on memory capacity and bandwidth. PATEO cites Thor's accelerated compute and high memory bandwidth as what supports responsive local inference within automotive safety permissions.

The cabin scenarios PATEO describes — fatigue detection prompting a window or seat massage, air-quality suggestions, parking-violation and flood alerts — point to an agent that acts on voice, visual and sensor inputs rather than only answering queries. Whether that lands in showrooms depends on the nominated program reaching mass production in the second quarter of 2027.