Either llama4:400b or qwen3:235b-a22b. There's limited gains above 100b-ish though. You might be better off running multiple smaller models as a team. Qwen3:30b-a3b with a 128k context is impressive