The people I know using this model are perpetually hamstrung by slowness, pedantry, misplaced assumptions and appeals to authority. You can't move forward if you're just trying to norm your training data.
They've created this extremely slow, overwrought system with lots of guardrails and token limits designed only to create addiction. Why would you use that when you can have a marginally less "smart" system that's 50x faster and 100x cheaper? Who cares if the model one shots something. It needs to be checked anyway, by test in a machine or a lab.
Who are the pedantic, authority-appealing users who are using this model?
What makes you believe they’re norming their training data to that group?
No US company will ever buy from Z.ai or DeepSeek - or the ones that do, will never admit to it.
Hence all the money has flowed to the top despite viable alternatives*
* there is a separate market of workloads contracted out to Fireworks, Baseten et al. that might be closer to what you're talking about. But it's still not going to touch them because those customers are a completely different buyer archetype