What changed
A fully crossed study of three models, five inference frameworks, six benchmarks, and four generation modes found significant score changes even under greedy decoding; backend choice explained roughly 39% of ordinary out-of-box variability.
Why it matters
A benchmark result is partly a property of the serving stack and defaults, not only the model weights and prompt.
What to do
Record framework, version, kernels, tokenizer, chat template, and every generation setting, then rerun acceptance tests whenever the serving stack changes.
Caveat
Only three instruction-tuned models were tested, so the 39% estimate is not universal and larger production-serving studies are still needed.