> GPT-NL was never built to compete with Claude or ChatGPT. It was trained exclusively on licensed data, and is intended more for governments and companies where privacy and compliance matter more than raw performance.”

That's it? That it didn't aim to compete with SOTA models? Maybe this is something you have to start with something, then ramp up, rather do what only a select few labs been able to do, start with really big models. Especially if you're resource constrained, which since this is a government project, I really hope for the sake of the tax payers it was.