“In an internal evaluation of a coding agent, a QIN3 base model had better latency and cost performance compared to GPT models.”