Something’s off.An article reports that the same Qwen2.5-1.5B-Instruct model file, given identical token IDs and decoded greedily, produced different written answers on an Apple M1 MacBook and an NVIDIA Jetson Orin Nano Super. In the reported 250-problem GSM8K math subset, 38 generated responses differed and 18 final answers differed, despite fp16 results whose overall scores were close.
The article attributes the split to tiny, deterministic floating-point differences between hardware and software kernels. It says a near-tie between candidate tokens can send a generated solution down a different path. Its int8 Jetson setup reportedly changed still more answers, but that setup also changed both weight precision and runtime, so the article says it did not isolate which change caused its larger score drop. The tests cover one model, one device pair and fixed benchmark subsets.
UNVERIFIED
Hacker News metadata is a discovery lead, not verification of the linked claim.
This source-attributed report has not been independently verified.
