DeepSeek launches V4.1 Flash: 552B asymmetric CED architecture slashes KV cache to 1/437
量子位 · wechat · 2026-09-10
DeepSeek 上线 V4.1 Flash，552B 参数 MoE，采用全新 Causal-Encoder-Decoder 非对称结构：输入激活仅 8B、输出 16B，多数 prompt token 只需走前 20 层编码器，prefill 计算量接近减半。这也是其首个原生支持视觉理解的多模态正式版模型，9 月 14 日起 V4.1 Pro 请求将被重路由至 Flash。
推理成本是最大亮点：CSA2 稀疏注意力沿条目、序列、层三个维度压缩缓存，主 KV Cache 降到 FP4，配合 SWA Bounded Replay 分层持久化，HBM 需求减至 1/4、SSD 减至 1/8——从 V1 到 V4.1 Flash KV Cache 总共缩小到原来的 1/437，核心目标是让长上下文 Agent 负载在经济上成立。
More from Infra
- Anthropic's noncancelable compute commitments hit $430B, per S-1 filing — Hesamation · 2026-10-27
- AI chip startup Tenstorrent raises $1.5B at $6.2B valuation, more than doubling — Beth_Kindig · 2026-10-27
- Anthropic discloses up to $120.5B SpaceX compute deal, ~$3.1B monthly at full capacity — XFreeze · 2026-10-27
- Satire: swapping Raft for an LLM-style decision model tanks write throughput 99.98% — irombie · 2026-10-27
- Cadence joins the domain-specific AI model camp for chip design — karlfreund · 2026-10-27
- Qdrant builds Universal I/O to run vector search mmap code on object storage — qdrant_engine · 2026-10-27