Scaling the quality curve to 70B models

Determine whether the quality curve of Leech-lattice vector quantization, whose direction has been measured on Qwen3-4B, Qwen3-8B, and Qwen3-14B without establishing a scaling law, reaches the 70B model class.

Background

The paper evaluates the same 2-bit $\Lambda_{24}(12)$ quantization configuration on Qwen3-4B, Qwen3-8B, and Qwen3-14B. Across these three sizes, perplexity degradation and the MMLU deficit decrease with model size, but the authors explicitly decline to infer a scaling law because the evidence comprises only three models and is sensitive to calibration draws.

The 70B class is the regime motivating the paper’s memory argument, but it lies outside the measured model sizes. The paper therefore leaves unresolved whether the observed improvement in quality with increasing model size continues at 70B.