Read this ad-free on Signal Reads → I deleted 400GB of Llama weights this morning. All of them. After three years of swearing by Meta’s ecosystem as the gold standard for local inference, a single 48-hour benchmark run on the Qwen 3.5 family made me realize I’ve been