Local Models Handle Half of Everyday Tasks, with Faster Response Times Than the Cloud
Localmaxxing
AI BriefingA five-week experiment found that, among approximately 1,400 everyday work tasks in total, about 50% could be successfully handled by a local 35B-parameter model, such as Qwen 3.6 35B, covering categories including email, scheduling, summarization, and administrative tasks. Performance comparisons showed that the local model had an average response time of 2.8 seconds on routine agent tasks, making it 2.1 times faster than the cloud-based Claude Opus 4.5, although the latter still led by approximately 20% on complex reasoning. Local model outputs were more concise, while cloud models were better in structure and code quality. As local model performance improves, shifting computing loads locally to address cloud costs has become an inevitable trend.
Source: Tomasz Tunguz blog · tomtunguz.com