CITIC Securities: DeepSEEK-v4.1-Flash Release Focuses on Three Main Investment Lines
The Zhitong Finance App learned that CITIC Securities released a research report stating that DeepSeek-v4.1-Flash was released to improve coding and agent capabilities while strengthening multi-modal understanding capabilities. The model innovation uses the CED architecture, combined with CSA2 cross-layer cache sharing, index multiplexing, and FP4 quantization, reducing the global KV cache to about 1/4 of V4-Flash. Architectural optimization freed up room for inference cost reduction, and DeepSeek announced that it will continue to lower API prices. Domestic models will continue to compete along the two paths of high cost performance and cutting-edge intelligence. Under DeepSeek's deep adaptation, domestic computing power infrastructure will benefit simultaneously with domestic models. At the same time, cost-effective reasoning will accelerate the implementation of FDE and enterprise-level agents.
CITIC Securities's main views are as follows:
DeepSEEK-V4.1-FLASH: Using the new CED architecture, the model capability surpasses DeepSEEK-v4-Pro.
On September 10, 2026, DeepSeek released DeepSEEK-v4.1-Flash, which uses a new “causal encoder-decoder” (CED) architecture, has multi-modal understanding capabilities, supports millions of token contexts, and simultaneously opens APIs and model weights. The backbone parameters of the model are 552 billion, and each token activates 8 billion and 16 billion parameters respectively during the pre-filling and decoding stages, further reducing inference costs for long contexts and agent scenarios. According to DeepSeek's official WeChat account, considering that V4.1-Flash has completely surpassed V4-Pro in various indicators such as performance, cost, speed, and total usage time, DeepSeek plans to launch the v4 Pro model in an orderly manner, but at this stage, considering that V4-Pro still has a large call volume and the time required for the Infra migration, V4-Pro is still maintaining service status as of September 15.
Programming and agent capabilities have been significantly improved, and multi-modal understanding expands the boundaries of task execution.
According to DeepSeek's WeChat account, 1) In terms of coding ability, V4.1-Flash reached 90.6 points in Terminal-Bench 2.1, higher than Opus 5's 89.1 points and GPT-5.6 Sol's 88.8 points. In Terminal-Bench 4.0 where the difficulty level was further increased, V4.1-Flash scored 31.2 points, lower than GPT-6 Astra's 57.9 points and Fable 5.1's 55.8 points; The DeepSWE v1.1 test reached 74.2 points, which is basically the same as GPT-6 Astra's 74.1 score, and higher than Fable 5.1's 67.4.
2) In terms of agent ability, AutomationBench and Agent's Last Exam scored 54.8 points and 31.8 points respectively, both higher than Opus-5's 50.3 points, 28.6 points, and GPT-5.6 Sol's 45.8 and 26.7 points. This shows that after improving visual understanding, the ability to perform general automation tasks and complex tasks has improved.
Architectural innovation: The CED architecture collaborates with CSA2 to optimize computational paths and cache reuse, continuing the ultimate cost reduction of long-term context inference.
1) Innovative “causal encoder-decoder” (CED) architecture, reducing global KV cache usage to about 1/4 of V4-Flash. The model's 40-layer Transformer network is divided into the first 20 layer causal encoder (Causal-Encoder) and the latter 20 layer decoder (Decoder). The output of the last layer of the encoder is directly projected to generate the global KV required for the decoder. There is no need to rely on the hidden state of each layer of the decoder to be generated layer by layer, so most input tokens only need to complete the encoder calculation. In order to preserve partial information, the decoder additionally processes 128 tokens at the end of the prompt, roughly reconstructing each layer's sliding window attention (SWA) cache, which nearly halves the amount of pre-filled computation in long input scenarios. Referring to the official DeepSeek technical report, the activation parameters for each token during the prefill (prefill) stage were reduced from 16 billion to 8 billion during the decoding phase. In terms of cache accuracy, the model reduced the global main KV cache from the previous generation FP8 to FP4, and introduced quantitative sensing training control accuracy loss in post-training, reducing this part of the storage usage by nearly half, while SWA KV retained FP8. The model architecture innovation combined with CSA2's cross-layer cache sharing mechanism reduced the global KV cache usage to about 1/4 of V4-Flash, simultaneously driving the new model's demand for HBM to 1/4 of V4-Flash and 1/8 for SSD.
2) CSA2 synchronously reduces cache usage and floating-point computation by sharing kV and multiplexing top-k token search results in a hierarchical manner. The first two layers of the model only use SWA, and the rest of the layers use compressed sparse attention CSA2 and are configured with Full, Reindex, and Reuse modes: in full mode, the model generates global primary KV and index keys through computation and filters the top-K position; Reindex shares the KV in the previous level in full mode, but recalculates the correlation to update the Top-K token; Reuse shares KV and Top-K results at the same time, eliminating repeated cache generation and index calculations. Specifically, in addition to layer 2 of SWA, the remaining 18 layers of the encoder are configured with “1 layer Full+5 layer Reuse” according to each set of 6 layers; the 20 layers of the decoder are divided into 5 groups, the first group is “1 layer Full+3 layer Reuse”, and the last 4 groups are “1 layer Reindex+3 layer Reuse”, so that the last 20 layers share the global KV generated by the first full layer. The above design combines reducing KV Cache with reducing Indexer search calculations. According to DeepSeek's official technical paper, when the model context is extended from 4K to 1M, the single-token decoding FLOPs only increased by about 25%, showing a significant improvement in long-context reasoning efficiency.
Reasoning optimization is transmitted to the API price reduction to maintain the cost-effective advantage of the domestic open source model.
According to DeepSeek's official pricing, V4.1-Flash's idle cache hit input, and output prices are 0.02 yuan, 1 yuan, and 4 yuan respectively, down 60.0%, 33.3%, and 11.1% from the previous generation of 0.05 yuan, 1.5 yuan, and 4.5 yuan respectively. The price during peak periods is double that of idle time.
Application impact: The new model continues the direction of extreme cost reduction, which favors the implementation of FDE and enterprise-level agent applications.
The model continues to reduce call prices while improving programming, agent, and multi-modal capabilities. It is expected to reduce the cost burden of multi-round tool calls and long-term context processing for enterprise-level agents and accelerate the implementation of FDE. At the same time, professional software companies that have accumulated industry knowledge, are deeply embedded in enterprise business processes, or have vertical private data barriers, and can deliver reliable results in strong regulatory scenarios are expected to take the lead in transforming model capabilities into product value and commercialization increases.
Investment strategy: It is recommended to focus on the following three main investment lines.
1) AI infrastructure: DeepSeek is deeply adapted to domestic computing power, and domestic computing power is moving in the opposite direction from the domestic model.
2) AI application: The model continues the open source strategy, further improves efficiency and reduces costs, and benefits FDE and software application companies with barriers.
3) Original model factory: On the one hand, architectural innovation continues to free up space to improve reasoning efficiency and reduce costs. DeepSeek-v4.1-Flash, GLM-5.3-Flash, and Qwen3.8-Flash, which were released in the past 2 months, have all significantly reduced long-term contextual inference overhead through architecture optimization, confirming that the architecture upgrade still has the potential to further reduce costs. On the other hand, domestic models will continue to compete around the two paths of high cost performance and cutting-edge intelligence. The Flash series covers daily programming, office, and general agent requirements with a lower number of activation parameters and call costs, while the flagship model with a large number of parameters is expected to continue to increase the upper limit of complex reasoning, expertise, and long-range task capabilities.
Risk factors:
The development and application expansion of AI core technology fell short of expectations; cost reduction of computing power fell short of expectations; serious social impact caused by improper use of AI; data security risks; information security risks; increased industry competition.