@frontenddogma@mas.to But it's still much weaker than current SOTA models. And the funniest part is that Mistral 4 tends to randomly drop Chinese words into long responses -- probably because they trained it on a ton of Chinese data due to the European copyright madness.