DeepSeek发布V4.1 Flash模型，推进KV缓存压缩技术

DeepSeek推出了V4.1 Flash模型，这是一个拥有552B参数的多模态混合专家模型，能够支持高达1百万tokens的上下文长度1。该模型采用Causal Encoder-Decoder架构、CSA2压缩机制和FP4量化等技术手段，在KV缓存存储方面取得显著进展1。

在性能指标上，V4.1 Flash的推理速度接近420 Tokens/s1，其中Prefill阶段激活参数为8B，Decode阶段激活参数为16B1。该模型将KV缓存存储相比V4-Flash版本压缩至运行时缓存的1/4和持久化缓存的1/8，实现了4倍的KV缓存压缩，全局KV缓存存储成本降低至890字节/token1。同时，该模型支持视觉输入功能，最大支持1024个视觉tokens，可处理约1344×1344像素的图像1。

DeepSeek has released DeepSeek-v4.1 Flash, a multimodal mixture-of-experts model with 552 billion parameters that supports a context window of up to 1 million tokens 1. The model achieves significant efficiency gains through advanced compression techniques, reducing KV cache storage to one-eighth of DeepSeek-v4 Flash while maintaining high-quality task performance 1.

The model employs a Causal Encoder-Decoder architecture combined with CSA2 compression and FP4 quantization to compress the KV cache 1. During the prefill phase, the model activates 8 billion parameters, while the decode phase uses 16 billion activated parameters 1. The global KV cache storage cost stands at 890 bytes per token, enabling inference speeds approaching 420 tokens per second 1. Additionally, DeepSeek-v4.1 Flash supports visual input with a maximum of 1,024 vision tokens, capable of processing images up to approximately 1,344 by 1,344 pixels 1.