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

CITIC Securities research notes state that since 2026, domestic large-model vendors have focused on upgrading Agent and coding capabilities, launching new models one after another. We believe the upcoming next-generation DeepSeek model is expected to continue the high cost-performance open-source model route, achieving stronger memory functions and handling ultra-long context, while refining coding and Agent capabilities and also addressing shortcomings in multimodal performance, bringing new investment opportunities in the directions of model providers, AI applications, and AI infrastructure.

1、Model providers: The next-generation DeepSeek model is expected to work together with other domestic models, driving China’s AI to accelerate toward the world. At the same time, model training will push cost reduction one step further, and cheaper tokens will drive an overall increase in global large-model API call volume. 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 thousands of industries and benefiting AI application companies with barriers to entry. 3、AI infrastructure: Cost reduction leads to increased usage, allowing AI Infra to benefit; domestic AI Infra and domestic models are moving toward each other.

The full text is as follows

Computer｜DeepSeek: Outlook for the Next-Generation Model

Since 2026, domestic large-model vendors have focused on upgrading Agent and coding capabilities, launching new models one after another. We believe the upcoming next-generation DeepSeek model is expected to continue the high cost-performance open-source model route, achieving stronger memory functions and handling ultra-long context, while refining coding and Agent capabilities and also addressing shortcomings in multimodal performance, bringing new investment opportunities in the directions of model providers, AI applications, and AI infrastructure.

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

In the AI programming domain, training framework upgrades, using complete code repositories and engineering trails as training data, and introducing deeper reasoning chains with multi-step execution and self-repair have produced AI Coding that 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 Agent managers who help AI deliver maximum effectiveness. In the multi-Agent cluster domain, 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 released “lobster-like” products, releasing the productivity of digital employees. In the native multimodality domain, native multimodal architectures have become the mainstream direction; hybrid embedding encoding has made rapid breakthroughs, but domestic models still need to make breakthroughs in key areas such as real-time audio-video interaction and cross-modal continuous reasoning.

▍ Domestic large models: Dense iteration and upgrades, continuous capability breakthroughs.

1）MiniMax: Coding capabilities are further upgraded. In the M2.7 SWE-Pro test, the score is 56.22%, surpassing Gemini 3.1 Pro; in the VIBE-Pro test for end-to-end full project delivery scenarios, the score is 55.6%, comparable to Claude Opus 4.6, and understanding of the operating logic of software systems has been further enhanced. At the same time, the M2 series models participate in training processes such as RL to implement self-iteration during the M2.7 training.

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 reconstruction, and deep debugging with very limited manual intervention. Its tool-calling and multi-step task execution capability (MCP-Atlas 67.8%), and its联网检索与信息理解 capability (Browse Comp 89.7%) are close to or even exceed those of overseas leading models.

3）Kimi: Kimi 2.5 introduces visual capabilities to automatically break down interaction logic and reproduce code. It also launches a new Agent cluster mode. In agent application test sets such as HLE-Full, BrowseComp, and DeepSearchQA, it achieves scores comparable to GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro. Moonshoot adopts a price-reduction strategy, with API prices reduced by more than 30% versus K2 Turbo’s pricing.

4）Xiaomi: The Xiaomi MiMo-V2-Pro is close to or even ahead of some overseas top models in test sets measuring model Agent calling capability, such as ClawEval and t2-bench. Its early internal test version, using the anonymous codename Hunter Alpha, was launched on OpenRouter. During the launch period, it topped the daily call-volume leaderboard for multiple days. We are optimistic that the large-model foundation will empower Xiaomi across the entire ecosystem of people, cars, and home, enabling a leap in AI capabilities.

▍ DeepSeek outlook: Continue the high cost-performance route, refine long-text, coding, 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 decreases by 60%/75%, respectively. Meanwhile, scores in coding and multi-Agent capability BenchMark increase significantly. Combined with DeepSeek’s model evolution direction and the Engram module paper with梁文峰 as a signing participant, we believe that next-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 computation of the attention layers in the Transformer architecture, thereby enabling ultra-long context handling. This improves model efficiency while refining coding and Agent capabilities and addressing multimodal shortcomings.

▍ Risk factors:

Development of AI core technologies and expansion of applications do not meet expectations, cost reduction in compute not meeting expectations, serious social impacts caused by improper use of AI, data security risks, information security risks, and intensifying industry competition.

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

1）Model providers: The next-generation DeepSeek model is expected to work together with other domestic models, driving China’s AI to accelerate toward the world. At the same time, model training will further reduce costs, and cheaper tokens will drive an overall increase in global large-model API call volume.

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 thousands of industries, and benefiting AI application companies with barriers to entry;

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

（Source: First Financial）