I don't know enough about Grok's architecture to understand how it could be sped up. We know that the Chinese models do all kinds of funky stuff to speed up, and I'd test qwen 3.8 to get a better feel for it, if I didn't get 499'd^[1] on it every other query - hopefully someone will fix something some day.

I feel like there are a lot of things made in China lately that are usable conceptually, especially for the now-runner-ups like Xai and Meta. And maybe all you really need is a RL'd MoE version of Grok for coding - they did try that with grok-coder, but that was the old team. I'm sure that the expertise at Cursor, and their tracing captures, can help with this.

1. This is very disruptive in a coding agent environment because you don't know if you can retry it... whomever thought it was a good idea to ever respond w/ 499 - needs to do a serious year of penance in my to-be-developed isolation chambers because that's the dumbest brain fart I have ever encountered - it used to be an internal nginx thing - how on earth can I see it on my client if the error is that the client broke connection before the server could respond? it's not possible to get a status code without a connection. ↩