Li Auto's OTA 8.6 brings Mach VLA 2.0 ADAS to 2022 cars across two chip generations

量子位 · wechat · 2026-09-23

Li Auto pushed OTA 8.6 to nearly a million AD Max owners (2022-2025 models), bringing Mach VLA 2.0 to three platforms at once: its in-house M100, NVIDIA Orin-X and Thor-U.

On the road: faster car-following, more decisive overtaking and lane changes, better negotiation in congestion; parking gains unlabeled-slot recognition, remote wake-up, and control via the Livis AI glasses ('park for me' from outside the car).

The hard part: M100 uses a dynamic data-flow architecture while Orin-X/Thor-U are GPUs with different quantization paths (FP4/FP8 vs INT8), so model structure, token budget, operators and deployment all had to be redesigned. The student model isn't trained to copy teacher answers but to drive itself, with the teacher correcting its hesitation and mistakes.

Why it matters: AI models iterate in months while cars live five-plus years. Li Auto's five in-house stacks — models, chips, StarRing OS, Harness, compilers — turn a model upgrade into a cross-chip capability-transfer engineering effort, potentially rewriting how automaker technical strength and product lifecycles are judged.

Original post →

More from Embodied

Embodied channel →