Qwen 3.5-2B recorded 69.4%... the baseline model is 21.6%

Qwen 3.5-2B's accuracy on the GSM8K math benchmark jumped to 69.4% after researchers modified a subset of its weights, up from 21.6% for the unmodified baseline model under identical conditions, according to a preprint posted to arXiv on November 10 by QDeep and ITMO University's LEAD team. The evaluation used 1,319 GSM8K problems in thinking mode with greedy decoding and an 8,192-token answer cap, a setting chosen after the baseline struggled most on the first 200 test problems; on the remaining 1,119 problems the tuned model scored 69.6% versus 21.6% for the baseline. The result matters because it suggests targeted weight edits, rather than full retraining, can sharply lift reasoning performance in small open models. What to watch next is whether the gains hold beyond GSM8K and how often responses hit the token limit in the full evaluation.