Calibration Language Changes How Quantized Models Fail: A Case Study of a Small Bangla Math LLM

Quantization lets large language models (LLMs) run on low-cost hardware, often the only option for speakers of languages such as Bangla. Yet quantization is calibrated almost only on English, and its effect on other languages is poorly understood, although Bangla needs 2.9x more tokens per word, giving quantization error more steps to accumulate. We ask how quantization, and the language of its calibration data, affect a small Bangla math model. We quantize Gemma-2 2B, fine-tuned on Bangla math, with round-to-nearest (RTN), NF4 and GPTQ at 8 to 3 bits, calibrate GPTQ on five sets including the same problems in Bangla and in English, and test 200 held-out problems with corrected paired tests. 8-bit RTN is lossless, 3-bit RTN solves nothing, and GPTQ beats RTN at 4 and 3 bits. Calibration language does not significantly change accuracy, but English calibration raises the share of repetitive, unfinished outputs by about 10 points (p < 0.001), even on identical problems, while changing only the domain has no effect. Conceptually, quantization damage in a low-resource language appears as repetition rather than language drift, and accuracy alone misses it. Practically, deployments should calibrate in the target language and monitor repetition.