CITIC Securities: DeepSeek's next-generation new model is expected to continue the high cost-performance open-source model approach

CITIC Securities research reports point out that since 2026, domestic large-model vendors have focused on upgrading Agent and coding capabilities, competing to release new models. We believe that the upcoming DeepSeek next-generation model is expected to continue the high-cost-performance open-source model path, achieving stronger memory capabilities and handling of ultra-long context windows, while further refining code and Agent capabilities and also addressing the shortfalls in multimodal performance. This will bring new investment opportunities across model manufacturers, AI applications, and AI infrastructure.

1、Model manufacturers: DeepSeek’s next-generation model is expected to work alongside other domestic models to help China’s AI accelerate toward the world. At the same time, model training will further reduce costs; cheaper tokens will drive an overall increase in global large-model API usage. 2、AI applications: Model parity helps ease market anxiety caused by narratives about conflicts between models and applications, supporting the deployment of AI Agents across industries, benefiting AI application companies with strong barriers. 3、AI infrastructure: Cost reduction leads to increased usage, allowing AI Infra to benefit; domestic AI infrastructure and domestic models are moving toward each other.

Full text below

Computers｜DeepSeek: Outlook on the Next Generation of Models

Since 2026, domestic large-model vendors have focused on upgrading Agent and coding capabilities, competing to release new models. We believe that the upcoming DeepSeek next-generation model is expected to continue the high-cost-performance open-source model path, achieving stronger memory capabilities and handling of ultra-long context windows, while further refining code and Agent capabilities and also addressing the shortfalls in multimodal performance. This will bring new investment opportunities across model manufacturers, AI applications, and AI infrastructure.

▍ Code, Agents, native multimodality: The upgrade direction for global large models.

In AI programming, with upgrades to training frameworks, adopting complete code repositories and engineering trails as training data, and introducing deeper chains of thought with multi-step execution and self-repair, AI Coding has moved from code-completion tools to project-level autonomous intelligent agents. Harness Engineer is expected to enable technical personnel to transition from being code engineers to becoming Agent managers who get the maximum performance out of AI. In the multi-Agent cluster area, the flagship product OpenClaw fully demonstrates the potential of multi-Agent systems. Domestic vendors such as Zhipu, MiniMax, Tencent, and Kimi have all launched “lobster-like” products, releasing the productivity of digital employees. In native multimodality, native multimodal architectures have become the mainstream direction; hybrid embedding encoding has quickly broken through. However, domestic models still need breakthroughs in key areas such as real-time audio/video interaction and cross-modal continuous reasoning.

▍ Domestic large models: Dense iteration and upgrades, with capabilities continuing to break through.

1）MiniMax: Coding capabilities are further upgraded. In the M2.7 SWE-Pro test, the score was 56.22%, surpassing Gemini 3.1 Pro. In the VIBE-Pro test scenario for end-to-end complete project delivery, the score was 55.6%, on par with Claude Opus 4.6, further strengthening understanding of the operating logic of software systems. Meanwhile, the M2 series models participated in training processes such as RL to enable M2.7 self-iteration.

2）Zhipu: GLM-5 introduces DSA and its own “Slime” architecture. With minimal human intervention, it can autonomously complete system engineering tasks such as long-horizon agentic planning and execution, backend reconstruction, and deep debugging. Its capabilities in tool calling and multi-step task execution (MCP-Atlas 67.8%), as well as联网检索与信息理解 (Browse Comp 89.7%), are close to or even exceed the level of overseas leading models.

3）Kimi: Kimi 2.5 introduces visual capabilities to automatically break down interaction logic and reproduce code. It has launched a new Agent cluster mode. In intelligent agent application test sets such as HLE-Full, BrowseComp, and DeepSearchQA, it scored in the range to benchmark against GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro. With a price-reduction strategy, Moonshoot’s API price is more than 30% lower than K2 Turbo’s pricing.

4）Xiaomi: Xiaomi MiMo-V2-Pro is close to and even ahead of some overseas leading top models in test sets measuring model Agent calling capabilities, such as ClawEval and t2-bench. Its early internal test version, launched under an anonymous code name Hunter Alpha on OpenRouter, topped the daily call-volume leaderboard for multiple days during the rollout. We are optimistic that the large-model foundation will empower Xiaomi across its “people, cars, and home” full ecosystem, enabling a leap in AI capabilities.

▍DeepSeek Outlook: Continuing the high-cost-performance route, improving long-text, code, Agent, and multimodal capabilities.

DeepSeek’s DeepSeek V3.2 released in January 26 adopts a sparse attention (DSA) + mixture of experts (MoE) architecture to improve training and inference efficiency and reduce costs. The input/output token pricing drops by 60%/75%, respectively. At the same time, its scores on code and multi-Agent capability BenchMark are significantly improved. Combined with DeepSeek’s model evolution direction and the Engram module paper co-authored with Liang Wenfeng, we believe that new-generation models such as DeepSeek V4.0 are expected to integrate Engram into the already mature DSA+MoE architecture. By using layered storage to store key frequently used information, the attention layer computation in the Transformer architecture can be reduced exponentially, thereby enabling ultra-long context processing. While improving model efficiency, it further refines code and Agent capabilities and makes up for the shortcomings in multimodal performance.

▍ Risk factors:

AI core technology development and application expansion not meeting expectations; cost reductions in compute not meeting expectations; severe social impact caused by improper use of AI; data security risks; information security risks; intensifying industry competition.

▍ Investment strategy: We recommend focusing on the following three investment main lines.

1）Model manufacturers: DeepSeek’s next-generation model is expected to work alongside other domestic models to help China’s AI accelerate toward the world. At the same time, model training will further reduce costs; cheaper tokens will drive an overall increase in global large-model API usage.

2）AI applications: Model parity helps ease market anxiety caused by narratives about conflicts between models and applications, supporting the deployment of AI Agents across industries, benefiting AI application companies with strong barriers;

3）AI infrastructure: Cost reduction leads to increased usage, allowing AI Infra to benefit; domestic AI infrastructure and domestic models are moving toward each other.

（Source: Yicai Finance）