DeepSeek Retires Its 1.6T Model for a 552B One

The title of the technical report tells you what this model is: it is about pushing the limits of KV cache compression. The KV cache is the temporary storage where a model holds what it has already read, and because it eats memory as the context grows, it governs both cost and speed. DeepSeek says it has shrunk that storage to 890 bytes per token. That is roughly a quarter of the previous generation, V4-Flash, and one 437th of the first generation, V1.

The way it was shrunk lies in cutting the model into two pieces. The company calls the structure a causal encoder-decoder, and it divided 40 layers into 20 encoder layers in front and 20 decoder layers behind. The company wrote that "the decoder's global KV cache is projected from the final encoder hidden state instead of coming from each decoder layer's own hidden state." Rather than each layer building up its own memory, the back half takes what the front half has organized and passes along.

The 51-page technical report METAL reviewed sets out the actual mechanism behind the smaller cache. It is a compressed sparse attention the company calls CSA2, and the report says it is split into three statically assigned modes. Full mode generates the global KV and performs the indexing that picks which tokens to attend to. Reindex mode reuses the global KV from a preceding layer but rescores with its own indexer query to select fresh top-K indices. Reuse mode takes both the global KV and the top-K indices from a preceding layer and runs sparse attention directly. In all three modes, the company wrote, each layer keeps its own global query and SWA KV. Of the 40 layers, only the first two encoder layers use sliding window attention; the rest are all CSA2.

The report also separates out where the savings actually land. The 890-byte global KV cache sits in HBM at all times, and that is about a quarter of V4-Flash's. On top of that, a deployment optimization called SWA Bounded Replay cuts the persistent KV cache, the one that lives on SSD or in host memory, to roughly an eighth of V4-Flash's. That eighth is the product of two things. The persistent cache no longer stores SWA KV, which nearly halves it, and the global KV that remains is compressed to a quarter. In the previous generation's deployment, SWA KV accounted for nearly half of the persistent cache's capacity.

The difference this design makes shows up in the number of active parameters. Only 8 billion per token switch on while the input is being read, and only 16 billion while the answer is being produced. The company explained that this asymmetry sharply improves cost efficiency on agent work, which is input-heavy. It accepts context up to 1 million tokens, and images were trained alongside text from the start, so it handles them directly.

Two more extensions sit on the compute side. Engram is a conditional memory module built to decouple memorization from computation, with 196 billion parameters split evenly across two modules. The idea is to handle recurring patterns by lookup rather than recomputing them through the network every time. DSpark is the speculative decoding piece. Its drafter runs three Transformer blocks with a 128-token sliding window, computing drafts for five positions in a single forward pass, while a confidence head predicts an acceptance probability for each position. The scheduler combines those estimates with profiled throughput curves to pick a verification length for each request on the fly.

In the comparison table on the model card that METAL reviewed, the places where the new model comes out ahead cluster on the agent side. On Terminal-Bench 2.1, which covers terminal work, it recorded 90.6, passing Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8. On DeepSWE v1.1, a code-fixing task, its 74.2 is almost level with Opus 5's 74.0, and on AutomationBench, which measures work automation, its 54.8 is ahead of Opus 5's 50.3. Its Codeforces rating of 3471 is higher than V4-Pro's 3348.

The same table also carries the places where it lost. On Terminal-Bench 4.0, which runs through multiple stages, its 31.2 falls more than 20 points short of Opus 5's 51.8, and on HLE, a collection of expert-level problems, its 36.8 leaves a wide gap against Opus 5's 56.3. On SimpleQA-Verified, which measures the knowledge held inside the model itself, its 42.3 is below V4-Pro's 55.2. With a backbone a third the size, the larger model comes out ahead on memorized knowledge.

The model card carries one more table that tells you how to read these numbers. It shows DeepSWE v1.1 run with the same model placed inside eight different agent harnesses, and the scores range from 65.6 to 74.2. The 74.2 DeepSeek used in its announcement is the mini-SWE figure; run through Claude Code it becomes 69.8. METAL reported earlier that DeepSeek had assembled a new harness team aimed at Claude Code, and DSH Minimal, the in-house harness that team built, came out at 90.6 on Terminal-Bench 2.1, above Claude Code's 88.0.

The company located the gains not in the algorithm but in the data. It said post-training followed the standard sequence of supervised learning, reinforcement learning and on-policy distillation without touching the algorithm. Instead it automatically synthesized agent tasks and execution environments at scale, expanding data, tasks and rollouts in stages. Pre-training used 45 trillion tokens of multimodal data, and sparse attention was trained at a length of 64,000 tokens before being extended to 1 million tokens at the 34 trillion token mark.

The schedule for retiring V4-Pro is written down to the date. From noon Beijing time on September 14, every request arriving at deepseek-v4-pro will be handled by the new model and billed at the new model's rates. The company wrote that "V4.1 Flash comprehensively surpasses V4-Pro in performance, cost, speed and total time taken." METAL reported earlier that V4-Pro launched formally with three adjustable levels of reasoning effort, and that flagship is now giving up its place after a single month.

DeepSeek is replacing its own flagship with a model a third the size, and it locates the basis for that in KV cache compression. On agent tasks there are several places where it comes out ahead of the top-tier models, but gaps remain on long reasoning and stored knowledge. Two things are worth watching. Whether the September 14 switch passes without a drop in performance on the user's side, and what standard the industry settles on for comparing agent scores that swing by more than eight points depending on the harness.