A small team just took a 27B open model and cut its hallucinations in half.
Built on Qwen, trained only on open-source environments, and it beats GPT-6 Luna on Terminal-Bench.
It also writes half the words to get the job done.
Their second version runs 7 long agent sessions at once on a single gaming card.
Free. Apache 2.0. Yours to run.
This is how local AI wins. I have been saying it for years, but I am a nerd.