50B+ MoEs with few active parameters, what's the sweet spot for intelligence, agent speed, and affordable fine-tuning?

I’m building a Polish General purpose legal Model that drafts documents, answers questions using legal sources, and has enough coding ability to handle some automation. The workflow is very tool-heavy: Question → many sequential tool calls → final answer/document Think Claude Code/Codex-style execut

中文全文翻译正在补齐，可先查看下方英文原文。

I’m building a Polish General purpose legal Model that drafts documents, answers questions using legal sources, and has enough coding ability to handle some automation. The workflow is very tool-heavy: Question → many sequential tool calls → final answer/document Think Claude Code/Codex-style execut