Over the past six months, a quiet but decisive shift has taken place on the Hugging Face Hub: Chinese AI labs have made "open source" the default, while American labs have made "open source" an increasingly cautious concession. The HF Summer 2026 report, published on August 14, supplies the numbers that justify this framing.
The numbers: licences have become a geography tag
The report tallies every open-weight model larger than 20B parameters released in the first seven months of 2026. Of the 178 Chinese models in that band, 59% ship under Apache 2.0 and 22% under MIT, together well over 80%, and almost none carry any non-commercial restriction. On the American side of the same size band, only 29% use Apache or MIT, 41% ship under custom terms, and the remaining 30% declare no licence at all. Pull a 20B-plus Chinese model from the Hub today and you can deploy it commercially by default. Pull an American model of the same size and you will, more likely than not, read three pages of legal text first.
What deserves equal attention is the denominator. 178 Chinese releases above 20B versus a much smaller American count. The denominator gap is itself the result of the rewrite of the open-source map. It is not that American labs do not want to open-source. Their volume, cadence and posture mean they simply cannot put the same number of 20B-plus models on the Hub.
Why the two sides diverge
Chinese labs follow what HF calls a "low-friction, high-coverage" route. The report names Moonshot, Moonshot-ai sibling labs, Xiaomi and Z.ai among those that barely release anything below 70B. A developer's first contact with them is a model too large to run on the hardware they own. But when they do open the weights, the licence is the most permissive available. DeepSeek and Z.ai ship their 700B-to-1.65T models under plain MIT, a generosity that no Western frontier lab is willing to match. The commercial logic is plain: weights do not earn money, APIs and cloud do. The wider the open-source coverage, the more downstream applications get built, and the larger the token volume routed back. The reported valuations of Z.ai and Kimi K3 are, in part, the market pricing of exactly this strategy.
The American camp runs a different playbook. NVIDIA's Nemotron 3 Ultra at 561B, Nemotron 3 Super at 124B, and Arcee AI's Trinity-Large at 399B are the rare American releases above 100B. All three come from companies that sell hardware or sell cloud. AMD and NVIDIA are the two organisations that published the most new open models in 2026. The report is explicit: hardware vendors have realised that open models are a way to sell chips. A model optimised for your hardware and freely available is the most direct proof that the hardware works. Meta's Muse Glimmer 30B revival belongs in the same bucket. The remaining American labs still posting above 100B either have a cloud behind them or silicon behind them. Pure-application American companies have largely stopped competing at this layer of the open-weight ladder.
The signal everyone missed: licences are tightening
Here is the part of the report that makes the timing of this article matter.
In the last few weeks, even some of the most permissive Chinese labs have started to tighten. The two newest and largest models, Kimi K3 and Qwen 3.8 2.4T, have begun adding non-commercial restrictions and revenue-share requirements to their licences. A reader on the HF comments pointed out that Kimi's terms bar companies above a certain annual revenue from serving the model without explicit authorisation from Moonshot. That is a different sentence from "MIT, do what you want".
The point is not "Kimi is no longer open." The weights are still on the Hub. The files still download. The model is still open-weight. The point is that the "free model + permissive licence" Chinese open-source formula has reached a commercial inflection. Once a model moves from being a piece of technical PR into being part of real business infrastructure, the licence stops being "free" and starts to have edges. HF's own summary lands the punchline: "the industry is likely to shift toward clearer monetization paths from open-source adoption."
So what
Open weights in 2026 are not charity, and licences are not decoration. They are a contract about how the model publisher gets paid after you ship the model. The Chinese camp uses low-friction licences as an acquisition tool aimed at API volume and ecosystem position. The American camp treats the licence as part of a moat aimed at preserving hardware, cloud, and a small number of vertical-model talking rights. Kimi K3 and Qwen 3.8 2.4T tightening their licences is not a counter-example. It is the natural downstream shape of the curve.
If you are a developer, the time you spend reading licence pages over the next year will quietly exceed the time you spend reading model cards.
(Primary source: Hugging Face blog, "State of Open Models: Summer 2026 Observations", https://huggingface.co/blog/state-of-open-models-summer-2026)