I've been using Gemma-4, both 26b (MoE) and 31b (dense) for a while. Both models are containers of normalized knowledge of all kinds of things, not just science or technology, but some of all of the internet. Gemma-4 is a free model from a company that makes money with Gemini. The output is almost always "encouraging", and rarely it formulates a critique, lists different approaches or formulates a competing view. It's sycophantic. Not as much as "Meta chat", or "Google AI" but the tendency is there.

Now I've started using Qwen3.8-27b (dense). The difference is stark: the reasoning style reminds me of some my Chinese colleagues: criticism and push-back, as a user with an idea I have real work to do, but it's easier to guide it to a specific dialectic in a Socratic dialogue.

Maybe Chinese companies, e.g., Alibaba, don't limit the usefulness of their free (open weights) LLMs.

If someone has some competence in a topic they appear to be much more useful than American models.

2/2