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 code capabilities, and are rolling out new models in a competitive race. We believe that the next-generation DeepSeek model that is about to be released is expected to continue the high-performance, low-cost open-source model route, achieving stronger memory functions and ultra-long context handling, while refining code and Agent capabilities and also addressing current shortcomings in multimodal performance, bringing new investment opportunities in the directions of model original manufacturers, AI applications, and AI infrastructure.

1、Model Original Manufacturers: The next-generation DeepSeek model is expected to work in tandem with other domestic models, driving China’s AI to accelerate toward the world. At the same time, by pushing model training forward another step to reduce costs, cheaper tokens are expected to drive an overall increase in global large-model API call volumes. 2、AI Applications: Model democratization helps ease market anxiety caused by narratives about conflicts between models and applications, supporting the deployment of AI Agents across industries, which is favorable for AI application companies with barriers to entry. 3、AI Infrastructure: Cost reduction brings growth in usage, benefiting AI Infra; domestic AI Infra and domestic models are moving toward each other.

Full text as follows

Computers | DeepSeek: Outlook for the Next-Generation Model

Since 2026, domestic large-model vendors have focused on upgrading Agent and code capabilities, and are rolling out new models in a competitive race. We believe that the next-generation DeepSeek model that is about to be released is expected to continue the high-performance, low-cost open-source model route, achieving stronger memory functions and ultra-long context handling, while refining code and Agent capabilities and also addressing current shortcomings in multimodal performance, bringing new investment opportunities in the directions of model original manufacturers, AI applications, and AI infrastructure.

▍ Code, Agents, native multimodality: Global large-model upgrade directions.

In AI programming, training framework upgrades, adopting complete code repositories and engineering traces as training data, and introducing deeper chains of thought with multi-step execution and self-repair have created a shift for AI Coding from code-completion tools to project-level autonomous intelligent agents. Harness Engineer is expected to enable technical personnel to transition from code engineers to Agent managers who help AI deliver maximum effectiveness. In the multi-Agent cluster area, the phenomenon-level product OpenClaw fully demonstrates the potential of multi-Agent systems; domestic vendors such as Zhipu, MiniMax, Tencent, Kimi, and others have all launched “crayfish-like” products, releasing productivity for digital employees. In the native multimodality area, native multimodal architectures have become the mainstream direction; hybrid embedding encoding has made rapid breakthroughs. However, domestic models still need to make further breakthroughs in key areas such as real-time audio/video interaction and cross-modal continuous reasoning.

▍ Domestic Large Models: Intensive iteration upgrades, with capabilities continuing to break through.

1）MiniMax: Code capability is further upgraded. In the M2.7 SWE-Pro test, the score was 56.22%, exceeding Gemini 3.1 Pro. In the VIBE-Pro test scenario for end-to-end complete project delivery, the score was 55.6%, comparable to Claude Opus 4.6; understanding of software system execution logic is further strengthened. Meanwhile, the M2 series models participate in training scenarios such as RL, enabling the M2.7 training process to let the model iterate on itself.

2）Zhipu: GLM-5 introduces DSA and its own “Slime” architecture. It can autonomously complete system-engineering tasks such as agentic long-horizon planning and execution, backend restructuring, and deep debugging with very minimal human intervention. In tool calling and multi-step task execution (MCP-Atlas 67.8%), as well as online retrieval and information understanding (Browse Comp 89.7%), its capabilities are close to or even exceed those of leading overseas models.

3）Kimi: Kimi 2.5 introduces vision capabilities to automatically decompose interaction logic, reproduce code, and newly launches an Agent cluster mode. In the intelligent agent application test sets such as HLE-Full, BrowseComp, and DeepSearchQA, it achieved scores comparable to GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro. Moonshoot uses a price-reduction strategy; the API price is lowered by more than 30% compared with the K2 Turbo pricing.

4）Xiaomi: Xiaomi MiMo-V2-Pro is close to or even ahead of parts of some overseas top models in test sets measuring model Agent calling capabilities such as ClawEval and t2-bench. Its early internal test version, using the anonymous code name Hunter Alpha, was launched on OpenRouter; during the rollout period, it topped the daily call-volume leaderboard for multiple days. We believe that the large-model foundation will empower Xiaomi across its people, vehicles, and home ecosystems, enabling a leap in AI capabilities.

▍ DeepSeek Outlook: Continue the high-performance, low-cost route; refine 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, improving training and inference efficiency and reducing costs. The input/output token pricing decreases by 60%/75% respectively. At the same time, code and multi-Agent capability BenchMark scores are significantly improved. Combining DeepSeek’s model evolution direction and the Engram module paper with Liang Wenfeng’s authorship, 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 for key frequently used information, it can achieve an exponential reduction in the computation load of the attention layers in the Transformer architecture, thereby enabling ultra-long context handling. While improving model efficiency, it will further refine code and Agent capabilities and make up for shortcomings in multimodal performance.

▍ Risk factors:

AI core technology development and application expansion do not meet expectations; cost reduction in compute power does not meet expectations; improper use of AI causes severe social impacts; data security risks; information security risks; intensifying industry competition.

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

1）Model Original Manufacturers: The next-generation DeepSeek model is expected to work in tandem with other domestic models, driving China’s AI to accelerate toward the world. At the same time, by pushing model training forward another step to reduce costs, cheaper tokens are expected to drive an overall increase in global large-model API call volumes.

2）AI Applications: Model democratization helps ease market anxiety caused by narratives about conflicts between models and applications, supporting the deployment of AI Agents across industries, which is favorable for AI application companies with barriers to entry;

3）AI Infrastructure: Cost reduction brings growth in usage, benefiting AI Infra; domestic AI Infra and domestic models are moving toward each other.

（Source: First Finance）