Nvidia’s research demonstrates that the architecture surrounding an AI model, including context, memory, and action management, can generate double-digit performance improvements while cutting token costs in half. Nvidia Labs released the open-source framework NOOA, which achieved scores of 82.2% on SWE-bench Verified and 85.1% on ARC-AGI-3 with GPT-5.6-sol, while keeping cost per test under 20 dollars. This approach based on traditional software engineering, including pass-by-reference memory management and elimination of context-compaction processes, halves token usage. By releasing this framework as open source, Nvidia aims to broaden its customer base beyond hyperscalers by increasing demand for its GPUs.
Source: Read the original article
