DeepSeek has released DeepSeek-V4.1-Flash, a 552B-parameter multimodal mixture-of-experts model whose central claim is not a benchmark score but a storage figure: 890 bytes of global key-value cache per token. That is roughly a quarter of what its predecessor needed at the same sequence length, and it is the number that decides how many concurrent agent sessions a given machine can hold. The weights and an accompanying technical report are on Hugging Face under an MIT licence.
Key takeaways
- DeepSeek-V4.1-Flash compresses its global KV cache to 890 bytes per token, about one quarter of DeepSeek-V4-Flash, with persistent cache falling to roughly one eighth.
- A Causal Encoder-Decoder split means only 8B parameters activate during prefill against 16B at decode, targeting agent workloads dominated by input rather than output.
- Reported scores include 90.9 on GPQA Diamond, 90.6% on Terminal-Bench 2.1 and a 3471 Codeforces rating, with a one-million-token context window.
Why KV cache became the bottleneck
Long-horizon agent workflows broke the old cost model. Each tool call extends the context, and the cache holding every previous key and value grows with it, occupying HBM on the accelerator and spilling into host memory and SSD when sessions are persisted for prefix reuse. Sparse attention already cut the arithmetic of long-sequence processing; what it did not cut was the storage and the bandwidth spent shuttling that storage around, which is where serving cost now concentrates.
DeepSeek's answer attacks the cache along three axes at once rather than optimising any single one. Head-count reduction shrinks each entry, block-level compression shrinks the number of entries, and cross-layer sharing removes duplicate copies entirely β the network retains just three encoder caches and one decoder cache across all forty layers. Layered on top, FP4 quantization in E2M1 format with per-16-channel scaling factors takes the remaining bytes down further.
What the encoder-decoder split actually does
The architectural change underneath is a Causal Encoder-Decoder, or CED, adapted from the You Only Cache Once line of work. The model's forty layers divide evenly: the lower twenty act as an encoder that builds reusable context representations, and the upper twenty derive their key-value entries by projecting the encoder's final hidden state rather than computing their own.
The consequence is that most positions in a long prompt only traverse half the network. Prefill activates 8B parameters per token while decode activates 16B, an asymmetry that matters precisely because agent traffic is input-heavy β a model that re-reads a large context and emits a short tool call spends most of its compute on the half that just got cheaper.
Attention inside the encoder uses what DeepSeek calls CSA2, combining a 128-token sliding window for local context with global sparse retrieval and cross-layer reuse. Each layer is assigned a static mode β Full, Reindex or Reuse β determining whether it computes fresh sparse-attention indices or inherits them, which avoids recomputing retrieval decisions forty times over. A separate deployment trick, bounded replay of the sliding-window cache, is what drops persistent storage to about an eighth of the previous generation.
How it scores
At maximum reasoning effort β the model exposes a 1-to-100 effort dial β DeepSeek reports 90.9 on GPQA Diamond, 74.1% on MMLU-Pro and a 3471 Codeforces rating. Agentic evaluations are the stronger suit, with 90.6% on Terminal-Bench 2.1, 88.1% on CyberGym and 74.2% of issues resolved on DeepSWE v1.1. Coding benchmarks land lower, at 79.4% on HumanEval and 60.6% on BigCodeBench.
Sparsity is aggressive: each layer carries one shared expert alongside 384 routed experts, of which six activate per token. The multimodal path adds a 32-layer vision transformer with patch size 14, capped at 1,024 visual tokens per image, and the model was pre-trained on 45 trillion tokens. An independent architectural breakdown by the engineer who writes as zartbot measured roughly 420 tokens per second in practice and argued the jump is large enough that the point-release name undersells it.
What it means for open-weight serving
The competitive framing here is about deployment economics rather than leaderboard position. An open-weights model under MIT terms that quarters its cache footprint changes what self-hosting costs, because memory and cache bandwidth β not raw FLOPs β are what cap concurrent sessions on a fixed GPU budget. DeepSeek has run this play before, pairing capable weights with unusually low serving cost when V4 Pro reached general availability at a fraction of frontier pricing.
The open question is quality drift. Cross-layer reuse and four-bit cache quantization are both lossy by construction, and aggregate benchmarks are poor instruments for detecting degradation that only surfaces deep into a million-token session. Whether 890 bytes per token holds up across long agent runs is something deployments, not evaluation suites, will settle.
FAQ
Is DeepSeek-V4.1-Flash open source?
The weights are published on Hugging Face under an MIT licence, which permits commercial use, modification and redistribution. As with most open-weight releases, the training data and full training pipeline are not included, so it is open weights rather than fully open source.
How much smaller is the KV cache than DeepSeek-V4-Flash?
Global KV cache lands at roughly 890 bytes per token, about a quarter of the previous model at equivalent sequence length. Persistent cache, which matters when sessions are stored for prefix reuse, drops further to around one eighth thanks to bounded replay of the sliding-window cache.
Why do prefill and decode activate different parameter counts?
The Causal Encoder-Decoder design lets the upper twenty layers borrow key-value entries projected from the encoder's output instead of computing their own. Long prompts therefore only need the first half of the network, activating 8B parameters, while token-by-token generation uses the full stack at 16B.
