Quoting Artificial Analysis

The gpt-oss-120b is presented as the most intelligent American open-weights model, offering significant efficiency benefits due to its smaller size. While its intelligence scores are behind models like DeepSeek R1 and Qwen3 235B, it is notable for its hardware efficiency. The 120B version is the most intelligent model that can operate on a single H100 GPU, and a smaller 20B version is the most capable model that can run on a consumer GPU.