Tencent Weaponizes Stem Algorithm to Shave Off Compute Latency in Large Language Models

Chinese technology conglomerate Tencent Holdings Ltd. announced a critical architectural breakthrough in large language model processing, highlighting Beijing’s ongoing corporate race to supercharge domestic artificial intelligence efficiency amidst strict Western hardware limits. The company’s specialized artificial intelligence infrastructure division unveiled “Stem,” a proprietary block-sparse attention algorithm engineered to streamline the intensive data processing phase that occurs when large language models parse massive, text-heavy documents. Tencent claims the mathematical framework reduces the time required to output an initial character by up to 3.6 times, executing long-text reasoning while draining just 25% of the computational power typically consumed by standard attention models….