For local model, my goal is keep context around ~128k to optimize capability vs speed without decoherence.

I'm up to ~7m context in a session without spiralling.

I think most of cloud models are like bloated windows.