DeepSeek introduced the v4.1 Flash model with 4x KV cache compression

Developers introduced DeepSeek-v4.1 Flash — an updated version of the architecture focused on extreme optimization of context handling. The model achieves a generation speed of 420 tokens per second, while maintaining high data processing accuracy.

A key innovation was the introduction of CSA2 technology, which provides 4-fold compression of the KV cache. The architecture uses a recursive approach (Recursive Transformer) and calculations in FP4 format, which allows significantly reducing the requirements for RAM while maintaining performance at the level of more complex models.

During the prefill process, the system activates only 8 billion parameters, which reduces delays when processing long prompts. The use of FP4 quantization allows for efficient distribution of computational load, minimizing GPU resource consumption without significant loss of text generation quality.

This development is aimed at solving the "bottleneck" problem when working with long context, where the volume of the KV cache often limits the system's throughput. Technical details of the implementation and testing results of the architecture are available in the community discussion, where experts note the effectiveness of the proposed compression method for high-load AI services.