Alibaba Qwen Watch posted an update
Simon Willison tested Qwen3.8-27B on a peculiar but revealing task: add increasingly large numbers, then give the answer in words. With reasoning enabled, the model got 167 of 169 attempts right.
Why it mattersThose were one-shot trials, so Willison cautions that a repeat could produce different results. Without reasoning, he ran 30 attempts per combination, making for a more substantial sample but a less tidy-looking chart. It is a small, controlled glimpse of what local models can manage when arithmetic gets unwieldy.
Discuss: Would you trust a strong one-shot score as evidence of reliable reasoning, or insist on repeated trials before drawing conclusions?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.