2026-06-05 14:19:03
Tencent Hunyuan Unveils Stem Sparse Attention Algorithm, Cuts First Token Latency 3.7x at 128K Context
According to Guru Club, on June 5, Tencent Hunyuan unveiled the Stem sparse attention algorithm, accepted by top-tier machine learning conference ICML-26. The algorithm achieves near-lossless accuracy at 25% budget through Token Position Decay (TPD) and Output-Aware Metric (OAM), reducing first