I have this same effect, Gemma 4 31 seems to handle it the best on the local models I've tried. The other problem I run into are run away responses where it runs through all 65k tokens before erroring out. It seems to happen mostly in scene transitions, starting or ending a scene.