State of Open Models: Summer 2026 Observations

In the AI world, time appears to move quickly. During the span of just a few months following the spring report of our biannual analysis, several key observations have emerged as of this summer. This report outlines those observations from January to August 2026 and provides the data supporting each one. Model and dataset repositories on the HF hub continue to expand on a daily basis.

Public model repositories grew from 2.43 million to 2.96 million, while datasets grew from 711,000 to 1 million, and Spaces expanded from 1.00 million to 1.44 million. The distribution remains skewed, with roughly 85.6% of models having fewer than 200 lifetime downloads, and 1.5% of repositories accounting for 99.2% of all downloads. This distribution remains unaltered throughout the observed period.

Historically, there was a clear progression path for labs in the AI world: they would begin by releasing smaller models and gradually progress towards larger ones. However, in 2026, several Chinese labs have skipped this progression entirely. In almost every month of 2026, the largest and most performant open model released by a Chinese lab has been larger than any model released by an American lab.

China's monthly ceiling for model parameters ranged between 754 billion and 2.78 trillion, while America's ceiling remained under 130 billion in five of seven months, with NVIDIA's Nemotron 3 Ultra reaching 561 billion in May and June, and Inkling from Thinking Machines Lab.

The lab landscape has bifurcated into two distinct groups. The first group, comprising Moonshot, MiniMax, Xiaomi, and Z.ai, typically releases models that exceed 70 billion parameters. This makes the first encounter with their models challenging for developers, as these models often surpass the capabilities of most developers' own hardware. The second group, including Tencent and Alibaba's Qwen, covers a broader range of parameter sizes, spanning from under 1 billion to well above that threshold.

Two factors have enabled the growth of the first camp. First, building large models is no longer a differentiating factor. Both Xiaomi and Meituan have released models with over a trillion parameters this year, and neither of these companies was a household name in open weights just a year prior. Second, a lab no longer needs to release a small model to attract attention, as the community's quantization layer enables the efficient execution of large models within days, a point we revisit later.

The size profile of models has transformed from a statement of capability to a statement of intent. A frontier-only portfolio stakes everything on benchmark position and API demand. Conversely, a full-spectrum portfolio aims to become the standard family of models for developers. Both strategies are rational, as they play for different prizes.

In the United States, participation in open-source AI remains growing, primarily driven by hardware vendors, AMD and NVIDIA. Both companies have released over 200 new model repositories each, surpassing all other entities. While the center of gravity has shifted, with Google and Meta ranking behind NVIDIA in new model releases despite their historical leadership, Meta's shift towards closed flagship models further underscores this change.