Study: 4-bit quantization cripples Gemma 4, Qwen 3.5 in low-resource languages

A new academic study from Mohammad Wathiq Soualhi, posted on arXiv and submitted for review at EMNLP 2026, investigates what the author calls the "multilingual quantization tax" — the performance degradation caused by 4-bit weight quantization when deploying Small Language Models (SLMs) on edge devices. The paper evaluates two frontier SLM families, Gemma 4 and Qwen 3.5, in a zero-shot setting across eight typologically diverse languages, using the MMLU ProX Lite and GlobalPIQA benchmarks.

The results reveal four distinct failure modes. First, "Typological Fragility": low-resource languages and specific non-Latin scripts experience representational collapse via architecture-specific double dissociations, sometimes failing to even produce valid task logits. Second, the "Home Language Fragility Paradox": a model's foundational pretraining pathways do not reliably shield its native language from precision-loss damage. Third, "Domain-Specific Forgetting": multi-step cross-lingual reasoning degrades significantly, while associative soft-science recall stays relatively robust. Fourth, "Quantization Resistance": highly saturated, typologically aligned domains resist deterministic degradation, with any post-quantization gains falling within statistical noise.

The work challenges the English-centric benchmarking common in quantization research. While 4-bit quantization remains essential for fitting SLMs on memory-constrained devices, this study shows that parameter truncation can disproportionately harm multilingual capabilities, especially for underrepresented scripts. The findings suggest that edge AI developers need to audit quantization effects on target languages, not just English, and that future compression techniques may need to be architecture- and language-aware to avoid silent structural collapse in real-world deployments.