45 days running Qwen 3.6 locally: great with docs, bad at guessing
Someone ran the Qwen 3.6 model locally as a for 45 days, upgrading their GPU from a single 3070 Ti to a 5070 Ti plus 4070 Super along the way. They compared it against their own testing setup across roughly 30 long , each running six-plus hours. The standout was , which felt the most reliable. , a version that only activates part of its parameters at a time, was okay but tended to ignore or forget instructions faster.
Gemma was tested for a couple of days and performed poorly — it couldn't even output proper characters in the tester's local (non-English) language, while Qwen, despite being a Chinese-made model, handled that language without issue. Qwen handled coding tasks and everyday problems well, including running a tool called OpenClaw (which reportedly does not work with the Hermes model). Given clear rules and the right information upfront, it performed reliably without having to guess. When tested on 30 questions about a design spec with the spec document provided as context, it answered all 30 correctly, citing real values.
Without the document, it got zero out of 30 right. In short, it reasons well when facts are available but is weak at guessing. It was also notably faster than Claude, which the tester says sometimes 'thinks' for over 15 minutes.
Key points
- was the most satisfying ; better instruction-following than the 35B A3B variant
- Gemma performed poorly at coding and failed to output proper non-English characters
- With a design doc as context, the model answered 30/30 spec questions correctly; without it, 0/30 — strong at fact-based reasoning, weak at guessing
- OpenClaw worked with Qwen but reportedly does not work with the Hermes model
- GPU was upgraded from a 3070 Ti to a 5070 Ti plus 4070 Super during the 45-day real-world test