Delangue: the real race may no longer be at the border

41% of Hugging Face downloads in spring 2026 for open Chinese models, the six most used models on OpenRouter all open. The center of gravity shifted while we were looking at frontier benchmarks.

In plain terms

The CEO of Hugging Face states that the competition that matters is no longer the one for the most powerful models, but the one for the models that people actually use. And those are becoming increasingly open - and increasingly Chinese.

The numbers, as of their date

Three metrics, reported by TechCrunch on July 14, 2026, point in the same direction:

• Chinese open-weight models accounted for 41% of downloads on Hugging Face in the spring of 2026;
• on OpenRouter, the six most popular models are all open - Tencent, Xiaomi, DeepSeek, MiniMax, Z.ai;
• on Vercel, open models handled nearly a third of AI queries in June.

Three platforms, three different user populations, same trend. That's what makes the signal interesting: it's not an artifact of measurement from a single aggregator.

Be careful, however, about what these numbers say exactly. A download on Hugging Face is not a request served in production, and the share on OpenRouter measures a population of developers sensitive to price, not the entire market. These three metrics describe adoption among practitioners who arbitrate on cost - not global inference volumes, where closed APIs remain massive.

Under the hood

What has changed technically is the useful gap. As long as the best open model was two notches behind in reasoning and tool-calling, the extra cost of the closed model was self-justified. When a GLM-5.2 becomes a credible alternative for a significant portion of tasks, the arbitration shifts: for an agent that loops thousands of iterations, the price per token is no longer a line item, it's the limiting factor of the product. "Sufficient" quality at a tenth of the price beats "frontier" quality - on the tasks where it suffices. The whole challenge is knowing which ones.

So what

Clem Delangue's argument is less technical than political: according to him, the main risk of AI would be the concentration of power - hence his defense of model transparency rather than restriction. Satya Nadella, at Microsoft, holds a similar discourse against lock-in by a single provider. Coming from the first partner of OpenAI, the position speaks volumes about how buyers now think about their dependence.

For those who build: the question is no longer "what is the best model?" but "what is my portability strategy?" A harness that cannot change models with an environment variable is a debt that the market will make you pay within the next twelve months.

Note - it's the exact counterpart of the thread we're following on the economic viability of the open frontier: these models are gaining usage. It remains to be seen who pays for their training.