CICC: DS V4.1 Flash unlocks hardware potential, domestic cards move toward the 10T model era.

CICC released a research report stating that the core significance of DeepSeek V4.1 Flash lies in achieving higher task completion efficiency and per-unit GPU output at lower compute, storage, and memory access costs through architecture and inference system optimization. Extrapolating from the optimization logic of V4.1 Flash, the bank sees that a 10T model originally requiring deployment on NVIDIA Blackwell-series 64-card systems could potentially be deployed on 64-card Hopper-series systems, meaning that domestically produced compute chips with mainstream performance in 2026-2027 (such as the Ascend 960 series) are expected to support the continued expansion of domestic models toward the 10T-20T parameter platform, once again confirming the closed loop of domestic compute demand logic. CICC's main points are as follows: Cost reduction: V4.1 Flash is not simply a smaller model, but a systematic reduction of redundant computation, storage, and data movement in long-context inference. CED splits a 40-layer Transformer into a 20-layer Encoder and a 20-layer Decoder, so that long prompts do not need to pass through all of the latter 20 Decoder layers, thereby reducing Prefill computation. CSA2 reuses Global KV and Top-K through the division of labor among three layer typesFull, Reindex, and Reuse; combined with FP4 Main KV and SWA Bounded Replay, Global KV Cache is reduced to about 1/4 of V4 Flash, and Persistent KV Cache to about 1/8. Meanwhile, Single-Pass mHC reduces repeated Activation reads and writes through Kernel Fusion, further alleviating HBM bandwidth pressure. Efficiency gains: V4.1 Flash achieves higher task completion efficiency with lower per-unit resource consumption through smaller dynamic computation, externalized memory, and higher Decode throughput. Its Prefill and Decode stages activate only about 8B and 16B parameters per Token, respectively; Engram strips some static knowledge out of the backbone's dynamic computation and retrieves it on demand; DSpark improves Decode efficiency through lightweight Draft, confidence prediction, and dynamic scheduling. Risks Model price competition risk, hardware supply chain risk, AI commercialization risk.