10B is smaller than I thought it would be, although it does make sense with how much RAM/VRAM (most likely shared RAM in the Air) I'd assume the car has. I wonder if there are plans to make a 32B+ model, maybe even 80B+ models that dedicated robotaxis or other future cars with beefier hardware could use? I would assume that this lets the model think through more complex scenarios and use the data provided to it (every sensor, camera, etc) better.
I'm also a bit confused about the decision to incorporate COT, given the fast reaction times needed for any autonomous driving system. Hopefully, it is "dynamic COT" as seen with some of Google and Qwen's text-based models, and it therefore isn't slowed down by thinking when it needs to react quickly, but thinks about things like a car 0.3 miles ahead on a highway.
I do wonder if I could incorporate this on my FTC robot for an autonomous driving demo! Would certainly be interesting to work with it, and of course it is on Hugging Face for me to use. Just gotta hook a beefy EGPU up with something like a Raspberry Pi, I suppose.