I could see them maybe using a custom version of Deepseek for that second tier. Could be a boon for them if the inference costs are a lot lower.