How DeepSeek V4.1 Flash Separates Reading and Generation

DeepSeek V4.1 Flash generates text one token at a time. A token is a piece of text, often a word or part of one. To choose what comes next, the model transforms the text already available through successive processing stages called layers. Its unusual feature is that most supplied text can take a…

DeepSeek V4.1 Flash generates text one token at a time. A token is a piece of text, often a word or part of one. To choose what comes next, the model transforms the text already available through successive processing stages called layers. Its unusual feature is that most supplied text can take a shorter path through those stages than a token being used to generate the next one.DeepSeek's technical report motivates this design with agents that repeatedly receive long tool results and accumulated history, making the cost of reading and retaining that information a major part of running the model. Why should preparing old text for later use require the same computation as deciding what to write next?The shorter route prepares memory for later use. Recent input still enters the upper half before generation, and each chosen answer token uses both halves.The design divides 40 layers into two halves. The first 20 prepare reusable memory of the input. Most input positions can stop there; a final slice of up to 128 positions also passes through the upper half to prepare its recent context. Once generation begins, each newly chosen token goes through both halves to predict the following token. The architecture is called a causal encoder-decoder, or CED. We will unpack those roles, then trace how sharing memory, selecting what to read, storing fewer bits, and rebuilding recent state change different costs.The useful distinction is already visible: preparing a record of earlier text and using that record to choose a new token need not follow identical paths. The shorter input route matters most when there is much more new text to read than text to generate.Table of contentsWhy one causal stack reads and writesWhy agent workloads change the budgetPrefill, decode, and the cache between themHow CED changes the old-token dependencyWhy generated tokens still use the full modelWhy this is a different encoder-decoderHow CSA2 shares global historyWhy sparse search needs a hierarchyHow the cache bytes add upWhat bounded replay rebuildsPutting the architecture togetherWhat the quality evidence establishesWhat this means for agent architectures Read more

Source: The AI Edge — Published October 7, 2026 at 15:00 UTC — Category: Research