Local models can handle 50% of real daily AI workload, and they do it 2x faster. That is the finding from a five-week self-study by Tomasz Tunguz, who logged 1,478 agentic tasks across seven categories including scheduling, email, summarization, engineering, and market research, then benchmarked Qwen 3.6 35B-A3B-4bit on a MacBook Pro M5 against Claude Opus 4.5 via API. Mean task completion: 2.8 seconds local, 5.8 seconds cloud.
The quality gap is real but bounded. Opus 4.5 scores roughly 20% higher on reasoning benchmarks and produces cleaner structure, better bullet points, and polisher code. Qwen outputs half the tokens and completes routine tasks correctly. For agentic pipelines where one model feeds output into another system, token brevity is not a consolation prize, it is an architectural advantage. The tricky part is in the read: both passed every task, but the engineering and market research categories split 50/50 between simple and complex, meaning the local model fails silently on the hard half.
Tunguz frames this as localmaxxing: the rational response to ballooning cloud inference costs, where users push routine workloads onto hardware they already own and depreciate. The argument is not about privacy or principle. It is about latency and arithmetic. If your laptop handles half your tasks at twice the speed, the calculus is immediate. The full piece is worth reading for the task taxonomy table alone, which gives anyone building agent workflows a concrete baseline for routing decisions.
[READ ORIGINAL →]