[noosphr]

I've found that model performance starts substantially degrading after 10%.

[Schlagbohrer]

I see with qwen3.6-35B running in Pi that it gets stuck in loops when the 262k context gets more than 30% full.