[Der_Einzige]

Long context generation is a sampling problem. Set your opencode to use a modern sampler like min_p or newer and you'll see models behave better at longer context.