Sorry to be pedantic, but I think it’s important not to call these models “open source”. Open weight is the equivalent of a compiled binary that can be run locally, so you’re basically getting Windows where it’s legal to reverse engineer, not Linux. The weights are permissively licensed, so the community is still free to quantize, fine-tune, reap, etc., which is analogous to something like Tiny11. Without the training data and the pipeline code, though, if Alibaba or Moonshot decide to go closed, the community won’t be able to fork their models and meaningfully continue development.

It’s better than what we are getting from “OpenAI” or Anthropic for the moment, but these are Chinese, for-profit companies that are no more altruistic than Google or Meta, and their open weight LLMs are just as likely to go closed as Llama or Gemma if the companies deem that to be in their best interest.