We evaluate two ways to reduce inference cost for small local language models—lowering weight precision and averaging over multiple answer‑option orderings—using Qwen3‑0.6B and Qwen3‑1.7B with simulated round‑to‑nearest quantization (8, 6, 4, 3 bits; fp16 as reference). Experiments run on 300 items from ARC‑Easy, ARC‑Challenge, MMLU and HellaSwag, measuring logits for options A‑D.\n\nSix‑bit quantization shows no accuracy loss against fp16, actually scoring +1.8 (range +1.0 to +2.5) and +0.9 ( +0.3 to +1.5) points higher across datasets, an effect we do not explain. Eight‑bit quantization drops performance by -0.7 (‑1.1 to ‑0.3) and -0.4 (‑0.7 to ‑0.2) points.\n\nAt 4 bits the answer‑slot skew rises from 0.15 to 0.53 (0.6B) and from 0.07 to 0.12 (1.7B). Adding Gaussian noise with the same weight SNR yields skews of 0.15 and 0.37 (two draws) and 0.19 (one draw), so it remains inconclusive whether rounding adds more skew than equal‑power noise. Flip rates increase but stay below a baseline of independent uniform errors matched for accuracy.\n\nAveraging four cyclic orderings recovers +12.6 points at 4 bits for the 0.6B model (and +5.3 points at fp16). We price bits and passes in a machine‑independent unit: weight bytes read multiplied by passes. In cost‑matched pairs, extra orderings beat extra bits only when weights have 6 or more bits, reducing the decision table to: use 6‑bit weights and spend the remaining budget on up to four single‑pass orderings.\n\nAll simulations use the simplest quantizer on two small models; a title‑level search found no prior work combining quantization with option‑order averaging.\n\nReview: For memory‑constrained local deployments, 6‑bit quantization together with a few order‑averaging passes offers the best accuracy‑cost trade‑off, especially for tasks sensitive to answer‑option ordering.