Where can you get 256GB of memory?

Really depends on the use case but one of the best all around models that’s multimodal, instruction tuned, has good tool support and json responses is gemma3-27b which you can run with about 48gb GPU mem.

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