模型與推論系統
DeepSeek V4.1 Flash Compresses the KV Cache With an Asymmetric Encoder-Decoder as Legacy Model Endpoints Are Rerouted
DeepSeek has released V4.1 Flash with native vision and a one-million-token context window. Using a Causal Encoder-Decoder, sparse attention, and an FP4 cache, it compresses the global KV cache to 890 bytes per token. The API has already replaced the previous Flash model, and starting September 14, V4 Pro will also be temporarily rerouted to the new model, requiring existing services to revalidate output and latency.
DeepSeek released V4.1 Flash on September 10. The central change is not simply a larger parameter count, but the division of its 40-layer Transformer into a 20-layer causal encoder and a 20-layer decoder. Instead of storing a complete set of input KVs separately for every layer, the decoder projects global KVs from the encoder’s final state, activating only about 8B parameters during prefill and 16B during decoding. The model also uses CSA2 to share KVs, index keys, and Top-K results across attention layers, while bounded replay reconstructs sliding-window state. The primary KV cache uses E2M1 FP4, with one E4M3 scale shared by every 16 channels. DeepSeek says the global cache has been reduced to 890 bytes per token—about one-quarter of the previous generation—while its persisted SSD footprint is roughly one-eighth as large. This is particularly important for long-history, multi-turn tool-using agents.
The model includes a 552B-parameter backbone, 196B parameters of Engram conditional memory, and DeepSeek-ViT trained from scratch. Its training corpus comprises 45T tokens, and its context window extends to one million tokens. This also explains the discrepancy between how different pages report the “model size”: the release post counts the 552B backbone, while the Hugging Face interface shows approximately 763B. Deployers should not treat the former as the size of the complete model weights. DeepSeek even says that large-scale self-hosting may require around 2,000 GPUs; releasing the weights under the MIT License does not mean ordinary teams can deploy the model inexpensively.
The API migration risk is more immediate: `deepseek-v4-flash` and the experimental vision endpoint are already being temporarily routed to V4.1 Flash. Beginning at noon on September 14, before V4.1 Pro goes live, `deepseek-v4-pro` will also be rerouted and billed at Flash pricing. Although DeepSeek reports substantial gains on agentic benchmarks such as Terminal-Bench 2.1 and DeepSWE, those results use the highest reasoning effort, a specific agent scaffold, and context windows of up to one million tokens; some comparisons also come from internal evaluations. Engineering teams should preserve regression samples from the previous version and separately measure time to first token, cache-hit costs, visual inputs, and tool-call formatting rather than assuming behavioral compatibility based solely on model aliases.