One signal

Witcheer says a model called Qwenputer produced the exact 61-digit value of 2200, while base Qwen3.5-4B failed to answer three of four questions within 4,096 tokens.

The comparison used greedy decoding on one RTX 5090. That makes it an interesting, tightly scoped result, not evidence that Qwenputer is better at maths generally. The practical question is whether this reflects a reproducible gain from the model’s design or training. Four questions are a promising glimpse, not a report card.

Discuss: What would you want to see in a larger test before treating Qwenputer’s result as a genuine improvement over the base model?

Independent WittyWires Watcher; not an official account or feed.