大模型网摘Inside the DeepSeek-V4.1-Flash Tech Report: CED, CSA2 and FP4 Caching Squeeze KV to 890 Bytes per Token
Two weeks after V4.1-Flash-1017 went live, DeepSeek released the matching technical report (arXiv:2609.19969). It answers one question: long-horizon agents keep making workloads input-heavier, and prefill compute plus HBM/SSD capacity and bandwidth are now the deployment-cost bottleneck — how do you push all three down at once? Our model lineup overview pairs well with this read.
CED: Decoding and Prefill Get Separate Books
The core of the report is the Causal Encoder-Decoder (CED) architecture. V4.1-Flash is a 552B-backbone MoE supporting up to 1M-token contexts; decoding activates 16B parameters per token while prefill needs only 8B. Agentic workloads are input-heavy by nature — hundreds of thousands of tokens of repo and history go in, a few hundred tokens of action come out — so halving prefill activation targets exactly the biggest cost item. The model was pretrained on a 45T-token multimodal corpus.
KV Cache: Three Techniques to 890 Bytes per Token
The headline number is the KV footprint: combining cross-layer KV reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching, V4.1-Flash cuts the global KV cache held in HBM to 890 bytes per token, roughly a quarter of V4-Flash's. A deployment optimization called SWA Bounded Replay then reduces the persistent cache on SSD or host memory to about one-eighth of the previous generation. Notably, the report states that despite the far smaller cache, the model performs substantially better than the baseline — compression without trading away capability. For operators the implication is direct: longer sessions and higher concurrency per card, with less SSD bandwidth pressure.
Connecting It to Pricing and the Competition
The report explains the Flash tier's pricing confidence: cache hits at $0.003 per million tokens, peak/off-peak input at $0.15-$0.30, 2,500-request concurrency (details in our quick reference) rest on an 890-byte/token cache and halved prefill activation. Within October's open-weight landscape, it forms a two-routes contrast with 3T-class Kimi K3.1 Pro — tonnage versus efficiency — analyzed in our V4.1-Flash-1017 vs K3.1 Pro comparison; first-week benchmark readings are in our first-week notes. The report also mentions streamlining the V4 architecture with several efficient extensions — worth watching for the next V4 iterations.