According to an ecosystem analysis published by Hugging Face, developers have built more than 151,000 derivatives on Alibaba's Qwen models. That figure leads every other model family and makes Qwen one of the largest foundations in the open model ecosystem.

What a derivative is

A derivative is a new model produced from another model's weights: fine-tuned, quantised, pruned or adapted to another language. It means someone downloaded that model and built something on top of it.

That measures something different from downloads. Downloads include curiosity; you can download a model, try it and drop it. Producing a derivative takes labour — gathering data, training, publishing the result.

Why Qwen

A high derivative count requires several things at once:

  • A permissive licence — it must be possible to build on it and republish.
  • A range of sizes — small models are cheap to fine-tune; a family with several sizes multiplies the experiments.
  • Sufficient base quality — there is no point deriving from a weak model.
  • Continuity — regular new releases keep an ecosystem alive.

Qwen offers all four together, and the number is the consequence.

What it signals

This data makes concrete a shift of weight in the open model space. The open-weight movement was for a period built largely on models published by Western labs; today the foundation the ecosystem builds on most comes from a China-based family.

That has a non-technical consequence. When a model family becomes an ecosystem's foundation, its design choices, the tendencies in its training data and its constraints propagate through thousands of derivatives. Whoever lays the foundation shapes much of what gets built on it.

The limits of the number

One caveat is needed in reading the figure: 151,000 derivatives does not mean 151,000 serious projects. The count includes automatically generated quantised builds, one-person experiments and abandoned repositories.

The comparison still holds, though: applying the same measure across all families, Qwen comes out in front. It is the position in the ranking rather than the absolute number that carries meaning.

Compared with closed models

This measure has no equivalent for closed models, which is exactly why the two worlds need measuring differently. A closed model's success is counted in usage volume, revenue and enterprise contracts. An open model's is counted in what gets built on top of it.

A closed model can reach millions of users while nobody derives anything from it; an open model may see less use while becoming the foundation of thousands of new models. The second weighs more heavily in shaping the industry's next period.

For smaller languages

A high derivative count has a local consequence too. Building models for smaller languages is feasible not by training from scratch but by adapting an existing model — which requires a permissive licence and a base of reasonable size.

That the Qwen family offers both explains why adaptations in many languages, including Turkish, are being built on it.