Post-Training Ternarization Compresses Qwen3-4B but Reduces Capability

Summary

An arXiv study finds that post-training ternarization can reduce Qwen3-4B storage substantially, but capability and perplexity worsen, while faster inference remains unproven.