“Open source models can get close to the performance of leading proprietary models but cannot leapfrog them due to a lack of sufficient compute resources.”