Are there any (small, ~10B) models that you would say are a good collaborator?

A user on reddit.com is seeking small language models, specifically around "~10B" parameters, that can act as good collaborators. They note that while many new "~30B" and "~10B" models score high on benchmarks, they often fail to challenge flawed premises. The user desires a model that "push back on dumb ideas enough" rather than being optimized for "one-shotting stuff," indicating a preference for critical feedback over simple task completion, especially from a "10B" model.