Microsoft has introduced a new caching technique for language models called 'You Only Cache Once' (YOCO), which optimizes memory usage by storing key-value pairs just once. This innovation aims to make large language models (LLMs) more efficient. Additionally, MemLong, another solution, allows LLMs to handle up to 80,000 tokens on a single GPU by storing past context in memory banks, significantly enhancing their processing capabilities by processing 20x more context. In China, Alibaba's Qwen and a new F5-TTS model have shown significant advancements in open AI. Furthermore, a Shanghai lab has released a 1 million context window open-source LLM, with future developments aiming for even larger context windows, potentially reaching up to 100 million or 1 billion tokens. Another method, KV Cache Compression, retains 97% performance by allocating cache budgets based on head importance.