Assistant Pepe 32B feels oddly human

The interesting part here is not raw intelligence, but behavioral tuning: taking a strong STEM-leaning base and making it sound less obedient, more abrasive, and more alive. If it keeps capability while changing tone, that is a real UX differentiator for chat, roleplay, and creative use cases.

• –Negativity bias is an explicit anti-sycophancy strategy, which could make the model feel more candid than standard assistant-tuned chatbots
• –Qwen3-32B is a hard base to steer away from structured, utility-first behavior, so the social style shift is the main story
• –“Human” in this context likely means messier conversational pacing, not literal empathy, which is useful for some users and off-putting for others
• –The 32B size makes this more interesting than a small novelty finetune: there is enough capacity for personality without immediately collapsing into nonsense
• –If the claims hold up in practice, this points to a broader lesson for model builders: style tuning can matter as much as benchmark chasing