According to a research report from BOC International, major AI models from both domestic and international players completed significant upgrades around the Spring Festival of 2026. As the capabilities of large models improve, two evolutionary paths—AI agents and multimodal applications—are becoming increasingly distinct. Concurrently, KNOWLEDGE ATLAS (02513) issued a price adjustment notice for its GLM Coding Plan before the holiday. This price increase signals that leading large model companies are not short of demand and is expected to drive sustained growth in demand for computing power hardware. Rising computing power costs have emerged as a new trend, highlighting supply bottlenecks. The computing power industry chain is anticipated to gain fresh growth momentum. Key viewpoints from BOC International are as follows:
Around the Spring Festival, major AI models, both domestic and international, underwent significant upgrades. On the domestic front, Chinese AI models received密集 updates: Moonshot AI released the K2.5 model on January 27; KNOWLEDGE ATLAS launched the GLM-5 model on February 12; MINIMAX-WP (00100) introduced the MiniMax M2.5 model on February 13; ByteDance released the Doubao 2.0 model on February 14; and Alibaba launched the Qwen 3.5 model on February 16. Internationally, major model providers also completed important updates during the same period. OpenAI released GPT-5.3-Codex in February, its first model specifically designed for real-time programming, capable of generating over 1000 tokens per second. Anthropic introduced the Claude Opus 4.6 and Claude Sonnet 4.6 models in February, supporting a 1 million token context window and better performance on tasks like agent planning. Google released the Gemini 3.1 Pro model in February, which doubled its reasoning performance.
As model capabilities advance, applications in AI agents and multimodal domains are entering a more mature phase. In the agent domain, an open-source tool named OpenClaw caused a significant stir in developer communities and major tech forums in early 2026. Evolving beyond a simple conversational chatbot, it has transformed into an "office partner" capable of taking over computers and assisting humans with specific office tasks, offering a preliminary glimpse into a major productivity transformation. In the multimodal space, ByteDance's SeeDance2.0 equips large models with production capabilities for the film and television industry, moving from "generating a clip" to "completing a work." The usability rate for generating a 15-second video increased from 20% to 90%, enhancing efficiency, reducing costs, and potentially propelling industries like animation and drama into a phase of scaled development. BOC International believes that as large model capabilities improve, these two evolutionary paths are becoming clearer: AI Agents, by understanding user intent and decomposing complex tasks while utilizing tools like MCP and Skills, are rapidly expanding into commercial office, legal, and financial applications; multimodal technology, through the generation of simulated data, is expected to be further applied to accelerate scenarios like embodied intelligence.
Rising computing power costs have emerged as a new sign, highlighting supply bottlenecks. The computing power industry chain is expected to continue benefiting. KNOWLEDGE ATLAS (02513) issued a price adjustment letter for its GLM Coding Plan on February 12. Adjustments included canceling the first-purchase discount while retaining quarterly and annual subscription discounts, and increasing package prices by 30% or more. According to the company's disclosure, this decision was based on factors such as growing market demand, an expanding user base, increased usage volume, and changes in actual usage and resource investment. This price increase indicates that leading large model companies face strong demand, which is expected to drive continued growth in demand for computing power hardware.
Risk factors include model upgrades falling short of expectations, domestic computing power supply failing to meet expectations, and slower-than-expected application adoption.
Comments