CITIC Securities: DeepSeek V4 and Multiple Model Updates, Computing Power Demand Remains Tight

On April 27, CITIC Securities reported that DeepSeek-V4 was released concurrently with GPT-5.5. The former, utilizing CSA/HCA attention compression and mHC network reconstruction, not only closely competes with proprietary flagship models but also reduces the cost of million-context inference to a very low level, successfully operating on domestically produced computing infrastructure. The latter relies on strong software-hardware coupling, focusing on high-complexity knowledge work, and continues the business model of high pricing matched with high intelligence. Meanwhile, the generational leap of the base model resonates with Agent frameworks like OpenClaw and Hermes, accelerating the commercialization of the Agent ecosystem through the expansion of intelligent limits and optimization of inference costs.