CITIC SEC: Large-scale model intensive iterative upgrade Focus on model original factory, AI application, AI infrastructure direction opportunities

CITIC SEC released a research report stating that since 2026, domestic large-scale model manufacturers have focused on upgrading agent and code capabilities, and have been competing to release new models. The upcoming next-generation model of DeepSeek is expected to continue the high cost-performance route of open source models, achieving stronger memory function and ultra-long context processing capabilities, improving code and agent capabilities, and filling in the gaps in multimodality. This brings new investment opportunities in the directions of model factories, AI applications, and AI infrastructure, and the firm recommends focusing on these three investment themes. 1) Model Factory: The next-generation model of DeepSeek is expected to work hand in hand with other domestic models, driving Chinese AI acceleration towards the world. At the same time, the training of models is further reduced in cost, and cheaper tokens drive an overall increase in global large-scale model API calls. 2) AI Applications: Model equity helps alleviate market anxiety caused by conflicting narratives between models and applications, helping thousands of AI agents in various industries land, benefiting AI application companies with barriers. 3) AI Infrastructure: Cost reduction leads to an increase in usage, benefiting AI infrastructure, as domestic AI infrastructure moves in parallel with domestic models. CITIC SEC's main points are as follows: Code, Agent, Native multimodal: The direction of upgrading global large models. In the field of AI programming, training frameworks upgrade, using complete code repositories and project trajectories as training data, introducing multi-step execution and self-repairing deeper thinking chains, creating AI Coding from code completion tools to project-level autonomous intelligent agents. The Harness Engineer is expected to shift technology professionals from code engineers to agent managers who maximize AI efficiency. In the field of multi-agent clusters, the phenomenal product OpenClaw fully demonstrates the potential of multi-agent systems, and domestic manufacturers such as KNOWLEDGE ATLAS, MiniMax, Tencent, and Kimi have all launched "lobster-like" products, unleashing the productivity of digital employees. In the field of native multimodal, the native multimodal architecture has become the mainstream direction, and breakthroughs in real-time audiovisual interactions and cross-modal continuous reasoning are still urgently needed in key areas for domestic models. Domestic large models: Intensive iterative upgrades, continuous capability breakthroughs. 1) MiniMax (00100): The code capability is further upgraded, with a M2.7 SWE-Pro test score of 56.22%, surpassing Gemini 3.1 Pro; in the VIBE-Pro test of end-to-end project delivery scenarios, it scored 55.6%, on par with Claude Opus 4.6, further enhancing the understanding of software system logic. At the same time, the M2 series models participate in M2.7 training in RL and other scenarios, achieving self-iteration of the model. 2) KNOWLEDGE ATLAS (02513): GLM-5 introduces DSA and independently develops the "Slime" architecture, able to autonomously complete long-range planning and execution, backend reconstruction tasks, and deep debugging with minimal human intervention, with capabilities in tool invocation, multi-step task execution (MCP-Atlas 67.8%), network retrieval, and information understanding (Browse Comp 89.7%) approaching or even surpassing leading overseas model levels. 3) Kimi: Kimi 2.5 introduces visual capability automatic disassembly of interactive logic, code replication, and new Agent cluster mode, achieving scores on applications such as HLE-Full, BrowseComp, DeepSearchQA close to or on par with GPT-5.2, Claude 4.5 Opus, Gemini 3 Pro, Moonshoot adopts a pricing strategy, with API prices reduced by more than 30% compared to K2 Turbo pricing. 4) Xiaomi (01810): Xiaomi MiMo-V2-Pro in tests measuring model agent calling capabilities such as ClawEval, t2-bench, approaches or leads some overseas leading models, with its early internal testing version launching OpenRouter under the alias Hunter Alpha, topping the daily call volume leaderboard for multiple days during the launch period. The firm is optimistic about the empowerment of large models to Xiaomi's holistic ecosystem, achieving a leap in AI capabilities. DeepSeek Outlook: Continuing the high cost-performance route, refining long text, code, agent, and multimodal capabilities. DeepSeek V3.2, released in January 2026, uses a sparse attention (DSA) + mixed expert (MoE) architecture to improve training and inference efficiency and reduce costs. The pricing of input/output tokens has decreased by 60%/75%, while the code and multi-agent capabilities Benchmark scores have significantly improved. Combining the evolutionary direction of DeepSeek models and the Engram module paper with Liang Wenfeng's participation, the firm believes that the next-generation models such as DeepSeek V4.0 are expected to integrate Engram into the mature DSA+MoE architecture, exponentially reducing the computational complexity of attention layers in Transformer architectures by storing critical commonly used information hierarchically, thereby achieving ultra-long context processing, improving model efficiency while enhancing code and agent capabilities, and filling the gaps in multimodal areas. Risk Factors: AI core technology development and application expansion fall short of expectations, cost reduction in computing power falls short of expectations, misuse of AI causing serious societal impacts, data security risks, information security risks, intensified industry competition.