Behind Trillions of Token Interactions Worldwide: Open-Source Models Move from "Dominance by One" to "Multipolar Competition," with Chinese AI Breaking into the Top Tier
```
Empirical research based on over 100 trillion tokens on the OpenRouter platform shows that the large language model market is undergoing a profound transformation. The share of open-source models has risen to 33%, completely breaking the monopoly of closed-source models. The market landscape has shifted from the dominance of DeepSeek to diverse competition, while Chinese open-source AI has risen strongly in this transformation, officially joining the world’s top tier.
On December 4, renowned Silicon Valley venture capital firm a16z and large model API platform OpenRouter jointly stated in a report that the core driving force of this transformation comes from the explosive growth of Chinese models. Data shows that the market share of open-source models developed in China soared from 1.2% at the end of 2024 to a peak of nearly 30% in mid-2025, with an annual average of 13.0%, almost equal to the 13.7% share of open-source models from the rest of the world. Chinese models such as Qwen, DeepSeek, and MoonshotAI, leveraging technical strengths and local adaptation advantages, have achieved a leap from marginal participants to core players.
The report indicates that the competitive structure within the open-source ecosystem is undergoing dramatic change simultaneously. After the "summer turning point" in mid-2025, the market rapidly shifted from DeepSeek family occupying over 50% share in a highly concentrated state to fragmented competition. By the end of 2025, no single model could continue to hold over 25% of the market share, and user selection logic shifted from locking onto the "best model" to flexibly combining 5-7 top-tier models.
In addition, the research report reveals several disruptive trends: medium-sized models (15B-70B) are replacing small models to become mainstream, agent reasoning capability has surpassed text generation as the core value, programming applications have surged from 11% to over 50%, and Asia’s market spending share has doubled from 13% to 31%. The rules of competition have shifted from leaderboard rankings to real-world usage retention and workload matching capability.
China’s Strength Reshapes the Open-Source Landscape
The report states that the open-source model market has formed a "dual-track structure" of "closed source defines the performance upper limit, and open source provides diverse value". By the end of 2025, the market share of open-source models steadily climbed to 33%, and this growth is not a short-term fad, but driven by continual iteration of high-quality models like DeepSeek V3 and Kimi K2.
The rise of Chinese open-source models has exceeded expectations. By the end of 2024, China’s model market share was only 1.2%. By mid-2025, its peak had reached nearly 30%. Chinese models such as Qwen, DeepSeek, and MoonshotAI demonstrate unique advantages in technical capability and local adaptation, signaling that Chinese AI has officially entered the global top tier of the open-source track.
From a global regional perspective, the overall rise of the Asian market is the most significant, with global spending share doubling from 13% in the early stage of research to 31%, becoming a key growth driver. Although North America remains the largest single region, its spending share has long remained below 50%.
Language distribution data shows that Simplified Chinese, with a 4.95% share, has become the second largest language after English, reflecting strong demand in the Chinese market.
From Monopoly to Multipolar Co-Governance
According to the report, by the end of 2024, the open-source market was highly concentrated, with the DeepSeek family’s V3 and R1 models accounting for over 50% of token usage, almost forming a "one-party dominance" situation. However, this pattern was completely overturned after the "summer turning point" in mid-2025.
With the intensive release of new models such as Qwen, Minimax, Kimi K2, and GPT-OSS series, the competitive barrier in the open-source market was broken. These new models achieved large-scale production-grade applications within weeks of release. By the end of 2025, no single model had sustained over 25% share in the open-source market.
User behavior patterns have fundamentally changed. Developers no longer default to locking onto the "best model", but now diversify combinations among 5-7 top-tier models. This marks that the open-source ecosystem has officially entered a sufficiently competitive "warlords divided" stage, and a multi-model ecosystem has become the industry norm.
"Medium is the New Small": Disrupting Size Perception
Empirical data from over 100 trillion tokens has completely overturned the traditional perception that the open-source ecosystem is dominated by small, lightweight models. Data shows that developers are reshaping the model size landscape through real-world actions.
Although the number of small models (<15B) continues to increase, their overall usage share keeps declining, and the market is highly fragmented, making it difficult to form stable user stickiness.
By contrast, medium-sized models (15B-70B) have experienced explosive growth from scratch, with mid-sized models like Qwen2.5 Coder 32B rapidly building a highly competitive ecosystem.
These models precisely match users’ needs for the "balance point of capability and efficiency", becoming the core growth hub of the open-source market and confirming the industry’s new consensus that "medium is the new small".
The large model (>70B) field also shows diverse competition, with models like Qwen3 235B and Z.AI GLM 4.5 becoming core benchmarks in testing, and users preferring flexible switching among multiple top-tier large models.
Chinese Characteristics in Application Scenarios
Looking at the overall task distribution of open-source models, role play, with over 50% token share, is the largest application, thanks to the natural advantage of fewer content restrictions in open-source models. Programming assistance ranks second at 15%-20% share, and its share is steadily rising.
But Chinese open-source models show significant differentiation. Unlike the global market where "role play dominates," programming and technology-related applications of Chinese open-source models account for 39% in total, surpassing the 33% share for role play.
This difference shows that Chinese open-source models already have the capability to directly compete with world-class models in productivity domains such as code generation and technical reasoning. Their value proposition tilts more toward professional efficiency improvement rather than entertainment interaction, which could open up unique competitive advantages for Chinese models in the enterprise market.
Agent Reasoning Leads Paradigm Shift
The most disruptive finding revealed by the research is a fundamental paradigm shift in LLM usage—from single-turn text completion to agent reasoning workflows involving multiple steps and tool integration.
The number of tokens processed by models optimized for reasoning skyrocketed from almost negligible at the start of 2025 to over 50% of total usage. This change is driven by both supply and demand:
On the supply side, the release of models like GPT-5 and Claude 4.5 greatly enhanced the upper limit of reasoning ability; on the demand side, users increasingly favor models capable of managing task status, following multi-step logic, and supporting agent-style workflows.
Accompanying the rise of agent reasoning are two key features:
Prompt length has increased dramatically: the average input tokens per request grew almost fourfold from 1.5K to over 6K, with programming task prompts exceeding 20K—three to four times higher than other categories;
Tool use is increasingly prevalent, with models like Claude 4.5 Sonnet and Grok Code Fast leading the way, marking the essential shift of LLMs from "text generators" to "action executors".
"Cinderella’s Glass Slipper Effect" Defines a New Moat
The research discovered a group of "foundational user cohorts" with super-high long-term retention, and proposed the "Cinderella’s Glass Slipper Effect" framework to explain this, defining the core moat of the AI era.
The core logic of the framework is: high-value "workloads" that remain unmet always exist in the market. Each new generation of models is a "glass slipper fitting" matching process; when a model first perfectly solves the technical and economic constraints of a specific workload, users will build workflows and data pipelines around that model, forming a very high switching cost and stickiness.
Data validates this logic: The early foundational user cohorts of Claude 4 Sonnet and Gemini 2.5 Pro had a 40% retention rate after five months, but for models like Llama 4 Maverick that did not achieve such matching, all user cohorts showed poor retention. Additionally, DeepSeek models also exhibit a unique "boomerang effect," where some lost users return after trying other models.
This finding shows that the real competitive barrier comes from the first matching of a "workload-model" and the subsequent highly sticky foundational user cohort—retention is far more critical than growth. The industry focus is shifting from minor leaderboard advantages to real-world empirical analysis and operational optimization, from competition between single models to flexible multi-model strategies, and open source and closed source, East and West, will continue to coexist and compete for the long term.
Risk Warning and DisclaimerThe market carries risks; investments must be made cautiously. This article does not constitute personal investment advice, nor does it consider the special investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are consistent with their specific circumstances. Investment according to the article is at your own risk. ```