CMSC International: Domestic model iteration significantly accelerates, domestic chip mass production in the second half of the year will benefit large model efficiency.

CMSC International released a research report stating that the iteration of domestic models before and after the World Artificial Intelligence Conference (WAIC) has significantly accelerated. Darkness of the Moon released Kimi K3, which has 28 trillion parameters and around 1 million tokens in context; BABA-W (09988) launched the preview version of Qwen3.8Max, with parameters reaching 24 trillion, and there are reports that DeepSeek is also about to launch the V4 official version. Previously, KNOWLEDGE ATLAS (02513) released GLM-5.2, MiniMax-W (00100) released M3 and pre-heated the unified multimodal generation model H3, TENCENT (00700) released the HY3 model, etc. The industry rating remains "recommended", and the main targets include: Alibaba's full-stack AI capabilities, Tencent's C-end product ecological advantages, KUAISHOU-W (01024) intelligent AI video generation model, KNOWLEDGE ATLAS programming model SOTA, and MiniMax native multimodal capabilities. At this WAIC conference, the industry noted that model manufacturers are focusing more on long-term tasks, Coding, Agentic, and multimodal directions. Overall, the leading model parameters have reached the level of 2 to 3 trillion, with capabilities comparable to Claude Opus 4.8, and maintaining cost efficiency and open source advantages. API pricing is significantly lower than that of leading models in the United States. In the upstream of the industrial chain, the competition focus of domestic computing power manufacturers has shifted from "single card parameters" to "system capabilities". Chinese chip manufacturers such as Huawei and Pingtouge have successively exhibited super nodes, integrating hundreds to thousands of chips into a unified computing power pool through high-speed interconnection, to make up for the performance gap of single cards through architectural innovation. The industry believes that the mass production of domestic chips in the second half of the year, such as Huawei's Ascend 950 series, will be a significant catalyst for the efficiency of large models, driving down inference costs and improving the gross profit margin of model companies. The report points out that the commercial potential of Agents remains undiminished, and AI products on the end side are entering a period of acceleration. Alibaba launched the Invisible Agentic Computer + Agent Native Cloud, providing 24/7 agent capabilities; Tencent promotes the AI Buddy Agent product matrix, and MiniMax launched the first fully processed Agent localized operating platform MiniMax Hub. Manufacturers continue to move Agents from chat assistants to independently executable enterprise organizational assets, with commercialization of Coding leading the way. Agent office/multi-agent governance/multi-modal are the focus of the next iteration. Overall, the competitive landscape at the model level is not yet defined, with major manufacturers still in a fast-paced iteration cycle, but there has been a differentiation in model routes and manufacturer positioning: trillion-level super flagship models continue to break through the upper limits of intelligence, tackling long-range reasoning, Agent Swarm parallel tasks and other high-level complex tasks; billion-level models focus on cost-effectiveness and deployment efficiency, achieving widespread availability in daily general tasks.