The connection drops twice before the download finishes. In a shared office in Lagos, a developer restarts it and turns back to the licence file open in a second window. A few pages of plain English: use it, change it, sell what you build with it.
He is pulling weights from a company in Hangzhou he will never pay, for a product whose users will never learn the model's name. A decade ago this ran the other way: an African team licensing software from a Western vendor on terms it could not negotiate. Now the terms arrive as a text file, and the vendor does not know he exists.
The inversion is worth taking seriously. China's leading labs give away models that cost a fortune to train, deliberately. The interesting question is not whether the gift is generous. It is what the giver expects back.
Key takeaways
- Alibaba's Qwen family logged about 2 billion downloads on the Hugging Face Hub in 2026, the clearest single measure of how far Chinese open weights have spread.
- Chinese models took 41% of downloads on Hugging Face in spring 2026, the largest national share, though downloads are not deployments.
- DeepSeek's V4.1-Flash lists at RMB 2 and RMB 8 (≈US$0.28 and US$1.13; HK$2.2 and HK$8.8) per million input and output tokens at peak rates.
- Releasing weights destroys the model layer's pricing power and pushes value toward serving, silicon and applications, where the same labs already sell.
- The State Council's AI Plus (人工智能+) programme treats open-source ecosystems as part of the build-out, so release is industrial policy as well as strategy.
Three reasons, only one of them noble
Start with distribution. Qwen logged about 2 billion downloads on the Hugging Face Hub in 2026, and Chinese models took 41% of the platform's downloads in spring 2026, ahead of US models (Hugging Face reports, 2026).
Downloads are a weak metric on their own. Their significance is what they leave behind: fine-tuning scripts for one architecture, quantised builds, tutorials, the chat template everyone's tooling expects. A family that occupies that layer is expensive to displace, because a rival must move a toolchain rather than win a benchmark.
Second, the squeeze. Open weights cap the price of capability. When a credible model is free to download and self-host, every closed model at the same level must justify a multiple. The spread is wide: Qwen3.7-Max at US$2.50 and US$7.50 per million input and output tokens (Alibaba Cloud), DeepSeek's V4.1-Flash at RMB 2 and RMB 8 (≈US$0.28 and US$1.13; HK$2.2 and HK$8.8) at peak rates, Doubao Lite at RMB 0.30 and RMB 0.60 (≈US$0.04 and US$0.08; HK$0.33 and HK$0.66) (company rate cards).
That makes the commodity tier uninvestable without scale, and pushes the releasing labs toward layers where a download does not set the price: serving, cloud, silicon, applications.
Third, people and defaults. Researchers choose ecosystems early and stay. A lab whose architecture appears in every tutorial recruits from a larger pool and sets the vocabulary — tokenizer, tool-call format, evaluation harness — that others must accommodate. The release is a talent strategy that looks like charity.
The bill is paid somewhere
Open weights are not open books. Training data, the full recipe and usually the reinforcement-learning details stay private; the licence covers the artefact, not the process.
The costs do not disappear. Training is paid once; serving is paid forever, by someone. If the flagship interface earns nothing, the company needs another line: cloud contracts from teams that downloaded the model and now need somewhere to run it, device sales, advertising, support.
This is why the giveaway and the compute build-out are the same story. The more freely the weights circulate, the more demand lands on the serving layer — financed in China at national scale, with AI Plus targets giving local governments a reason to standardise on models they can host themselves.
What would reverse it
Three things could close the tap.
- A licensing shock. If export controls were extended from chips to weights, or Chinese regulators restricted overseas access to the strongest models, the download would stop being free.
- A security incident. A fine-tune that strips the safety behaviour, deployed at scale and traced to the original family, turns distribution into liability.
- Serving costs outrunning cloud revenue. Huawei's Ascend 950DT carried an indicative price above RMB 250,000 (≈US$37,200 / HK$292,000) in September, 20%–50% above quotes from two months earlier (Reuters, 10 September 2026). If serving costs rise faster than the cloud revenue they attract, free distribution becomes an expense a lab cannot carry.
What the giveaway buys
The logic is easiest to see from the receiving end. A team in Lagos, Jakarta or São Paulo that builds on a Chinese architecture will eventually need inference at volume, a hosted endpoint, someone who can support it at short notice. The download is the start of a relationship.
According to Li An, chief scientist at BrainNet (脑机网), the release strategy is best read as an attempt to make Chinese architectures the default substrate for other people's products, not as a concession about capability.
That framing explains the odd behaviour of giving away the thing that looks most valuable. The model is not the product. The model is the invitation.
Honest limitations
Download counts measure curiosity and convenience, not production use; a model pulled onto a laptop and abandoned counts the same as one serving a business. The claims about Chinese families' share of top open models rest on Hugging Face counters and company statements. Training data, compute spending and the internal economics of each release are undisclosed, so any explanation for a specific release is inference. Licences move as well: a permissive text today is no guarantee for the next release.
What readers can do now
- Operators: treat the licence file as a procurement document. Check commercial terms, acceptable-use clauses and whether the grant survives fine-tuning.
- Investors: value the serving layer, not the model. Ask which cloud, which chip and which support contract captures the demand a free download creates.
- Researchers: record the exact checkpoint and licence you evaluated. A result tied to a model that can be re-licensed or withdrawn is not reproducible.