Frost & Sullivan, in collaboration with LeadLeo, released 'China GenAI Market Insights: A Panoramic Study on Enterprise-Level Large Model Calls, 2025'

The evolution trend of large models is showing a direction of 'open source as mainstream, lighter parameters, and more widespread applications'.

Open-source large models are rapidly reshaping the industrial landscape. In 2023, 149 basic models were released globally, of which 65.7% were open-source; by 2025, the number of open-source models will far exceed closed-source ones, becoming dominant in the ecosystem. Parameter lightweighting has reduced the computing power threshold, making model training and inference more efficient and flexible, accelerating penetration into vertical industries and small and medium-sized enterprises. Transparent and verifiable structures, global developer collaboration, and flexible customization capabilities strengthen the autonomy and controllability of open-source models while reducing costs.

Closed-source models still have advantages in high-performance training, centralized optimization, and closed-loop services. However, their high cost and exclusivity limit their widespread adoption. It is expected that over 80% of enterprises will adopt open-source large models in their intelligent construction in the future, and the open-source ecosystem will become the core force for technology popularization and digital transformation.

Source: Frost & Sullivan analysis, LeadLeo research institute

The average daily call volume of the large model market has shown explosive growth.

In the first half of 2025, the average daily usage of Chinese large model market exceeded 10 trillion tokens, a growth of about 363% compared to the second half of 2024, showing explosive growth in usage. The surge in calls has significantly increased computing power and storage consumption, marking that large models are moving out of the pilot phase and entering the stage of large-scale implementation.

Large models have become the core engine for enterprises' digital and intelligent upgrading. At the same time, they are driving the accelerated reconstruction of the industrial chain in terms of computing power supply, data governance, application development, and service delivery, evolving from point breakthroughs to system integration.

Source: Frost & Sullivan analysis, LeadLeo research institute

Open-source models have become a key driving force behind a new round of growth in the enterprise-level market for large models

The current market still predominantly deploys closed-source models, with closed-source models accounting for 56% and open-source models for 44% in the first half of 2025. However, in future new call intentions, open-source will account for 70% and closed-source for 30%. The market is evolving from 'closed-source pilots' to a hybrid architecture that is 'open-source-led with closed-source as a supplement'. Enterprises will adopt a model of 'open-source as the foundation, closed-source for enhancement': open-source will handle large-scale, low-cost traffic (such as search enhancement, structured extraction, batch processing), while closed-source will focus on high-value, high-security, and ecosystem-bound scenarios (such as financial compliance, cross-cloud collaboration, fully managed suites).

Source: Frost & Sullivan analysis, LeadLeo research institute

In enterprise-level application scenarios, "business value" is the core consideration for selecting open-source or closed-source models.

Enterprise model selection is always guided by 'business value'. The advantage of closed-source models lies in their 'convenience and reliability', occupying mainstream markets with stable performance, a mature ecosystem, and rapid iteration. Open-source models provide 'autonomous control', support for deep customization, integrate models with enterprise data and business logic, ensure intellectual property and data security, and solidify training results into enterprise-specific digital assets.

Source: Frost & Sullivan analysis, LeadLeo research institute

The Internet and financial industries have become the core scenarios with the most active application of large models, accounting for over 90%.

The call pattern of large models exhibits industry heterogeneity: the internet accounts for 72%, relying on content generation and intelligent interaction, forming a low-latency, high-concurrency mode; finance follows closely behind, but due to high compliance reasoning and interpretability requirements, it shows a low-frequency, high-value characteristic; industries such as manufacturing, automotive, and transportation mostly adopt embedded or phased calls, emphasizing local deployment and private domain customization, reflecting the 'low-frequency, high-value' pattern. Overall, the call structure evolves from a single high-frequency pattern to a diversified embedding pattern, forming a new pattern where internet finance coexists with traditional industries in a low-frequency, high-concurrency scenario.

Source: Frost & Sullivan analysis, LeadLeo research institute

The core value of current large models lies in "improving quality and efficiency," which is mainly reflected in scenarios such as question answering enhancement, document processing, code assistance, intelligent customer service, and marketing content creation.

The commercial applications of large models focus on 'improving quality and efficiency'. Among them, 'QA enhancement' accounts for the highest proportion (32%), including search-based and knowledge-base-enhanced Q&A services provided by enterprises to external end customers and internal workers. Other important use cases include code assistants, document generation, intelligent customer service, and marketing content creation. Large models are evolving from auxiliary tools to business support bases.

Source: Frost & Sullivan analysis, LeadLeo research institute

Multiple vendors are accelerating the implementation to shape a new ecosystem for Chinese large models

At a critical stage of accelerating intelligent enterprise evolution, Alibaba Cloud, Volcano Engine, DeepSeek, and other enterprises are becoming the core engines for enterprise-level large model calls, leveraging a continuously iterating technology system, diversified model offerings, and scenario adaptation capabilities.

With the prosperity of the open-source ecosystem and the diversification of call patterns, China's large model industry is accelerating its transformation from 'technology-driven' to 'application-led', entering a stage of scaled popularization. Multiple manufacturers are promoting industrial implementation through competition and collaboration, forming an open and robust new ecosystem.