QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization

Name: QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization
Uploaded: 2026-09-08T06:22:31.686927+00:00
Duration: 405 s
Description: The episode discusses a paper titled "QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization." The hosts explain how QTEA achieves high accuracy while compressing large language models using ternary quantization. Key findings include a 16.7% relative gain in zero-shot accuracy on Qwen3-14B and a 7.2x speedup over standard FP16 baselines.

The episode discusses a paper titled "QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization." The hosts explain how QTEA achieves high accuracy while compressing large language models using ternary quantization. Key findings include a 16.7% relative gain in zero-shot accuracy on Qwen3-14B and a 7.2x speedup over standard FP16 baselines.

Open the full episode