Google Research has released the TurboQuant quantization compression algorithm, which can compress the KV cache of large language models to 3 bits, reducing memory usage by 6x and improving computational speed by 8x. The algorithm performs excellently across multiple benchmark tests, aiming to address model cache bottlenecks, and will be presented at ICLR 2026.

Google Researchが発表したTurboQuantアルゴリズムは、大規模言語モデルのKVキャッシュを3ビットに圧縮でき、メモリ使用量を少なくとも6倍削減し、訓練なしで精度を維持します。このアルゴリズムは、PolarQuantとQJLの2つのサブアルゴリズムを用いて従来の量子化を最適化し、多くの長いコンテキストのベンチマークで優れた性能を示しています。

Gate News Alert: On March 13, according to Gate's market data, as of press time, LYN (Everlyn AI) is trading at $0.30, up 57.05% in the last 24 hours, with a high of $0.50 and a low of $0.15. The 24-hour trading volume reached $26 million. The current market cap is approximately $7.71 million, an increase of $2.8 million compared to yesterday. Everlyn AI is the first decentralized autonomous video AI infrastructure layer with the fastest video generation speed globally. The platform provides multiple AI generation functionalities including image-to-video, text-to-video, text-to-image, and image-to-image. Compared to industry standards, Everlyn AI delivers 15x faster generation speed, 25x lower costs, and 8x higher architectural efficiency.