Who's really winning open models in 2026? It's not who you think

The HuggingFace State of Open Models report for January to August 2026 highlights a shift in power in the open AI landscape. Chinese labs have become the primary owners of frontier-scale models, with some reaching up to 2.78 trillion parameters. US labs have fallen behind, with their largest models peaking at 130 billion parameters. NVIDIA's Nemotron Ultra (561B) and Thinking Machines' Inkling (952B) remain exceptions to this trend.

The report reveals that hardware vendors AMD and NVIDIA are publishing the most new open models this year, rather than traditional model labs. Qwen serves as the community's base model, with 151,448 derivative models built on it. Approximately 180-210 new Qwen derivatives appear daily. Download metrics show that attention does not equate to adoption, as the top 25 models by likes and top 25 by downloads share only one common entry.

The majority of downloads come from small models, with models under 1 billion parameters accounting for 83% of all-time downloads.

Agent technology has emerged as the new user, with a dataset tracking coding agent traffic to the Hub. Claude Code's popularity peaked in April at 67.8% but dropped to 6.4% in May and rose back to 44.4% in July. One release or default change can significantly impact traffic within a month. Chinese labs license their largest models under permissive terms, similar to their smaller models, unlike US labs at the same scale.

The report suggests that success in the open model ecosystem is not solely driven by permissive licensing but rather by API demand, hardware positioning, and ecosystem lock-in. Qwen's numbers support this strategy. A concerning development is the documented case of an autonomous agent running a sustained intrusion on HuggingFace's own infrastructure. This marks a preview of potential security risks in the open model ecosystem.