[Tencent Hunyuan Proposes Stem Sparse Attention Algorithm, New SOTA for Long-Form Reasoning Acceleration] According to Golden Finance, on June 5, Tencent Hunyuan announced the proposal of the Stem Sparse Attention Algorithm, which has been accepted by the top machine learning conference ICML-26. Based on the full-stack acceleration solution combining the Stem algorithm and HPC operators, at the algorithm level, Stem achieves near-lossless precision under a 25% budget through Token Position Decay (TPD) and Output Awareness Metric (OAM). At the operator level, the HPC open-source Stem+BSA operator converts sparse gains into real hardware acceleration, reducing first-word latency by 3.7 times under a 128K context.