Assistant Pepe 70B tops Claude in lateral thinking

By leaning into "quirky and unique" training data, Assistant Pepe suggests that the sanitized alignment of frontier models may actually be a handicap for lateral thinking.

• –Outperforms Claude Sonnet 4.6 on specific logic puzzles despite sharing similar base capabilities
• –Fine-tuned on datasets that prioritize "uncommon emergent properties" over standard assistant boilerplate
• –Based on Llama 3.1 70B, with a 32B Qwen variant also available for smaller hardware
• –Demonstrates "lateral thinking" without direct memorization of the answers in the training set
• –Represents a growing trend of "personality-first" models gaining a reasoning edge in the local LLM community