I’ve published a new article, “GPT-5.1 vs Ministral-3 3B: Evaluating AI Newsletter Quality For Local AI Newsletter Generation”, where I test whether a $0 local model can match a frontier cloud model for automated newsletters.

I ran the same 6-stage n8n workflow—news ingestion, relevance scoring, editorial decision, research enrichment, formatting, and delivery—with both GPT-5.1 and Ministral-3 3B (via Ollama on a MacBook). Using LLM-as-a-judge evaluation across five topics, the local model scored 3.47 vs 4.07—closer than expected. The post covers three techniques that make local models viable: task decomposition, JSON auto-correction, and short context windows.

You can read it on Akribic here.