# Survey Finds Open Models Lead Use but Trail Closed Models in Production

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/38291
Published: 2026-10-22T20:04:08.000Z
Updated: 2026-10-22T20:04:08.000Z
Section: Technology

> Qwen led cumulative model-file requests while DeepSeek led routed token volume; 50% of developers surveyed used both open and closed models.

Open models reached more developers than closed models in a May 2026 survey, but closed models were more likely to be in production, highlighting a gap between adoption and deployment.

The survey covered 1,494 developers adding AI features. Overall, 79% used open models and 71% used closed models; 50% used both. Among users of each type, 51% of open-model users reported production use, compared with 63% of closed-model users.

Download and token-traffic measures point to different leaders and capture separate parts of adoption. The figures should not be treated as interchangeable measures of market share.

A snapshot labeled March 2026 counted 942.1 million cumulative requests to Alibaba’s Qwen model files and 476 million to Meta’s Llama files. It includes data through February. Counts include automated requests to model-file endpoints, so they do not represent unique developers.

Organizations headquartered in China accounted for 1.15 billion of the requests, compared with 723 million for U.S.-headquartered organizations. The grouping follows publishers’ headquarters, not downloader locations.

On a routed AI platform, DeepSeek models processed 14.37 trillion tokens from Nov. 3, 2024, through Nov. 30, 2025. Qwen logged 5.59 trillion and Llama 3.96 trillion during the same period. Routed totals exclude direct use through services such as ChatGPT, Gemini and Doubao.

An Aug. 1-31, 2026, ranking on a routing platform placed open-weight models in eight of its 10 top positions by token volume; seven of those models were built in China.

For routed traffic from May through September 2025, open models were estimated to account for 20% of model-layer usage and 4% of revenue. The usage mix later changed, while the revenue share has not been recalculated.

Survey respondents most often pointed to infrastructure or compute costs as a barrier, at 27%. Security, privacy or compliance concerns followed at 26%, then maintenance and updates at 24%, and deployment, hosting or scaling complexity at 23%. Those barrier responses covered 1,411 people.

These measures use “open” in different ways. Open-weight refers to public availability of a model’s weights, while the survey and routed-traffic estimate use the broader term “open models.” Public weights alone do not establish that training information, code and licensing meet a strict open-source definition.