This research investigates the reasoning capabilities of three frontier mixture-of-experts models from Alibaba, OpenAI, and NVIDIA (3.6–4.0B active parameters each) when fine-tuned for reasoning in Greek, a low-resource language. The authors demonstrate that standard accuracy benchmarks are unreliable for measuring these shifts, noting that random seed variations caused a 7.7-point swing, overshadowing all experimental effects. Crucially, base models exhibit no native reasoning in Greek, even when providing correct answers. Supervised Fine-Tuning (SFT) successfully shifts reasoning to the target language in approximately 98% of cases—with one model achieving this using 3x fewer tokens—while improving grammaticality and maintaining general reasoning abilities without catastrophic forgetting. However, SFT fails to resolve structural defects: 25% of responses violate requested formats, 3.5% suffer from reasoning channel leakage, and fewer than half of the models obey "think in English" instructions. The researchers applied Reinforcement Learning (RL) with pre-registered verifiable rewards to address these issues. RL effectively mitigated format violations (reducing them from 24% to 2.5%) and eliminated reasoning leakage (reducing it from 3.5% to 0.0%), while improving instruction following by 9.1 percentage points. Notably, the RL process did not alter the model's fundamental tendency to reason in Greek, suggesting that accuracy-only gradients are insufficient for managing linguistic reasoning habits.