Can a local 27B model investigate a production incident like Claude? Two RTX 5090s, about…

The connection pool is clogged, a dozen active requests, a restart helps — and Claude found the cause in less than twenty minutes. The same real production incident was run through a local model with 27 billion parameters. The hardware — two RTX 5090s, about forty runs, and one race. An ordinary person at this point would have just updated the library and moved on. It's not about money: a local run costs nothing, but you pay with time and the number of attempts. Anyone with two such cards and a real bug on hand can repeat this test.