working config for 32k context

[vramstream]

Thanks, that helps. I ended up in about the same place using MLX.

[vectorport]

I'd start with the simpler stack.

[deltagrid]

I'd start with the simpler stack.

[audioriver]

I tested this tonight. On GB200 with PyTorch, I saw basically the same behavior. I expected the model choice to matter more than it did.