How about context length? Isn't it still just a few thousand tokens on local models?