Google's TurboQuant technology, released in a paper in March this year, compresses the KV cache by reducing data previously stored in 16 bits to 3 bits, minimizing information loss to maintain performance. Google explained that this reduces KV cache memory use by one-sixth. Nvidia's "Trientance" technology, detailed in a paper published in April after research with the Massachusetts Institute of Technology and China's Zhejiang University, filters data by importance to selectively use key information, reducing KV cache use by 10.7 times compared to the previous method according to benchmark results. China's DeepSeek unveiled a memory-efficient technique called "Engram" in a paper earlier this year, which leaves frequently used data outside the GPU and calls it only when necessary, allowing the GPU to focus on reasoning about new inputs.