DeepSeek-v4.1 Flash: Pushing the Limits of KV Cache Compression

The latest version of the DeepSeek-v4.1 model, a knowledge-based cache compression algorithm, has been released. The algorithm uses a combination of knowledge base and neural network-based methods to achieve state-of-the-art performance in compression ratios and decompression speed. The authors of the paper claim that the model can achieve a 3.5x improvement in compression ratio compared to the previous version, with a 1.5x improvement in decompression speed. The model has been trained on a dataset of 1.5 billion bytes and has been tested on a range of tasks, including image and text compression. The code for the model has been released under the Apache 2.0 license.