Researchers Scale Post-Training Ternarisation to Qwen3-8B

A research team has presented an end-to-end evaluation of scaling an aggressive post-training ternarisation pipeline from Qwen3-4B to Qwen3-8B. The pipeline combines KOTMS rotation, E2M-ATQ adaptive ternarisation, and GPTQ-style error compensation in a weight-only A16 configuration.

The 8B model achieved a three-corpus perplexity ratio of 1.361x across WikiText-2, C4, and PTB. On eight zero-shot tasks, it reached a mean accuracy of 64.6% compared to 72.4% for FP16, demonstrating 78.5% chance-corrected capability retention. This outperforms the matched 4B run, which retained 69.6%, highlighting an advantage in model robustness as scale increases.

The resulting packed checkpoint measures 8.24 GiB while preserving recorded perplexity. Direct packed execution demonstrated an inference speed of 15.52 tokens per second within 7.35 GiB of memory, though a preliminary packed GEMV implementation remains slower than FP16 cuBLAS.