Maximizing Local Model Usage

Classifying 1.4k daily tasks over the past 5 weeks showed that a 35B local model could handle roughly half of them.

The largest share was Other 521 (35.3%), followed by Scheduling 254 (17.2%), Market Research 192 (13.0%), Summarization 184 (12.4%), Email & Inbound 170 (11.5%), Engineering 147 (9.9%), and Admin 10 (0.7%).

Notably, Email & Inbound, Scheduling, Summarization, and Admin — 618 tasks (41.8%) — were sufficiently handled locally, while Market Research and Engineering split roughly evenly between simple and complex tasks. The conclusion is that half of daily tasks run fine without the cloud.

The reasons cited for local models were privacy, cost, and hardware depreciation, but the actually felt value was latency. Under the same prompt and warm-up conditions, Qwen 3.6 35B-A3B-4bit averaged 2.8 seconds on MacBook Pro M5, while Claude Opus 4.5 averaged 5.8 seconds via API — about 2.1x faster.

• Claude Opus 4.5 scored about 20% higher on reasoning benchmarks.
• Local models are still 3-4 months behind the frontier.
• However, Opus excelled at structure and polish while Qwen excelled at short outputs, and both models completed the tasks correctly.

In the end, the natural response to tokenmaxxing is localmaxxing. As local models get faster and the gap narrows, more workloads are likely to shift to personal hardware.