I tried to run a 27B coding model on my 16 GB Mac, ofc it failed. It crawled at 0.4 tokens per second, about 200 seconds per tool call. it sucked haha The same base model on @darkbloomai answered in about 2 seconds. So I ran an Experiment: can the models on Darkbloom's network of idle Macs be the brain of a real coding agent? The setup: six Darkbloom models, the six tools an agent needs (read, write, edit, shell, grep, glob), and coding tasks executed for real in a sandbox, with every result checked rigorously Results: • 60/60 tool calls parsed cleanly • 67/71 multi-step tasks solved: fixing a bug, renaming across files, finding a planted AWS key, patching a SQL injection • Qwen3.8 27B, Gemma 4 26B and GLM 5.3 Flash went a perfect 12/12 • Total cost of every run: about 7 freakin cents ( so cheap btw) The common view is that open models are fine for chat but too flaky for agents. I'm starting to think otherwise. Here's why it matters. Coding agents burn enormous numbers of tokens, and most of what they do is routine: read a file, search the repo, run the tests. My bet is that 25-40% of an agent's turns can run on @darkbloomai at a fraction of the price, with only the hard steps staying on frontier models. That's a whole new demand category for a network of idle Macs if successful. Experiment 2 tests exactly that on real traffic... out soon !!!