Cockpit AI Shouldn't Rush to Chat—First Fix the Underground Garage Dead Zone

Here's a thing. At the Beijing Auto Show, ModelBest upgraded its SuperMate edge-side cockpit solution again, focusing on three areas: full-scenario Agent, full-modality interaction, and multi-chip platform adaptation. They also had a real car on-site for people to sit in. I squeezed in and sat for about ten minutes, and came out with my head full of one question: does it work in the scenario I know best?

The one I know best is the underground garage. B3, no signal, dense pillars, narrow spaces, and only a few spots where you can turn around. After doing memory parking for a while you realize the place users really curse is at the breakpoints. You get out, walk a couple steps, come back — does it recognize the route from just now? You take over the wheel for a moment midway — can it pick back up along the way? These things are inherently unfriendly to cloud-based large models, because there's simply no network in the garage. So I buy the edge-side path — SuperMate putting the model on the car's computer is logically aimed right at this. There was one line in the promo I noted down.

How cockpit AI can intervene in the user's car life at the right time and in the right way.

Translated into product language, this is about trigger timing and state machines — much harder to do than the model itself. Edge-side also has edge-side costs.

It partners with platforms like Intel and Rockchip. The upside is covering cars at different price points; the downside is the same experience has to land on chips with quite different compute, and what actually comes out is most likely determined by the weakest chip. I understand the tradeoff, but users don't know what chip tiering means — they just know the car next door performs better in the garage than mine.

My judgment is one line: today's cockpit AI doesn't lack ones that can chat — it lacks the one that can still remember who you are and where you're going after the network drops. They say 300,000 vehicles installed by end of 2026 — sounds like a lot, but what I want to see more is how many of those 300,000 are cars under 150k.