Open-weight AI models from China have captured a record share of usage on a leading US web development platform, signaling a shift in developer preference towards more cost-effective and customizable AI solutions amid rising prices for proprietary models.
- Open-weight Chinese AI models accounted for 54% of token volume on Vercel’s platform.
- DeepSeek-V4-Flash became the top-used model, outpacing OpenAI’s GPT-5.6 Luna.
- High costs limit adoption of proprietary models like Anthropic’s Fable 5 despite cutting-edge performance.
What happened
Open-weight AI models originating from China have reached unprecedented popularity on the US-based Vercel AI Gateway platform. On a recent Tuesday, open models represented 54% of all token usage, surpassing proprietary AI models which accounted for 46%. The weekend before recorded an even higher peak of 62% usage for open-weight systems. This marks a sharp rise from late June when open models were a minority with just 28% of token volume.
This surge is largely powered by DeepSeek’s lightweight DeepSeek-V4-Flash model, which currently leads in token volume usage on the platform. Other top models are also Chinese-developed open-weight offerings including StepFun’s Step 3.7 Flash and Zhipu’s GLM-5.2. These collectively dominate usage, with OpenAI’s GPT-5.6 Luna ranking only third. The data reflects a clear preference by developers for cheaper, customizable AI suitable for demanding production environments.
Why it matters
The growing dominance of open-weight models signals a pivotal shift in AI development and deployment economics. As proprietary models like Anthropic’s pricey Fable 5 struggle to gain traction due to cost constraints, open models provide a more accessible and scalable option for developers, especially for complex tasks like autonomous agents that consume large numbers of tokens for reasoning and tool integration.
This democratization of AI usage expands innovation opportunities but also raises new security concerns. Industry experts have pointed to the customization potential of these models being exploited by state-linked hacking groups for automating malicious activities. Meanwhile, the overall AI market is seeing an increase in US investment aimed at competing with Chinese open-weight AI leaders, exemplified by Nvidia’s multibillion-dollar deal with startup Poolside to build rival models.
What to watch next
Industry observers should monitor how the balance between open-weight and proprietary AI models evolves, particularly as cost sensitivity by businesses influences adoption patterns. The upcoming Nvidia-backed Nemotron model and efforts to scale US-developed open-weight systems will be key tests of domestic competitiveness against Chinese offerings like DeepSeek and Moonshot AI’s Kimi K3.
Additionally, heightened scrutiny around the security risks posed by customizable open-weight AI models is expected to increase. How regulators, cloud platforms, and cybersecurity firms respond to potential misuse will shape the trajectory of open AI models’ market penetration and trustworthiness in critical applications.