@ben_burtenshaw
the wildest part of this intelligence per watt paper (71.3% of chat queries could be local) is that the model is only a gpt-oss 20b. which is about a year old! compared to the current batch of small moe models (gemma 4, liquid LFM, Qwen-3.6, etc.) this is nothing. https://t.co/d4Oem5d35t