Reddit user suggests 4B LLM with local Wikipedia over 122B

Argues any LLM, even a 3T-parameter model, will hallucinate world knowledge. Suggests a 4B model with a kiwix skill and local Wikipedia (~50GB) to avoid this, built via a coding agent with fallback to live internet knowledge.