Why China Is Giving Away Its Best AI Models

01 The Giveaway That Shouldn't Make Sense

The puzzle, stated plainly: at the same time the leading US laboratories are closing their best models behind paid interfaces, major Chinese laboratories are publishing models of comparable capability as files anyone can download. Not demo access, not a trial tier - the learned parameters themselves, free, at a moment when the technology is widely described as the most valuable of the decade. Giving away the crown jewel looks like a contradiction in terms.

It is not charity, and it is not carelessness. The behavior is consistent across several independent labs and several years, which is the signature of strategy rather than accident. When multiple competing organizations converge on the same unusual policy, the policy is usually paying for itself in some currency other than subscription revenue.

This article separates the layers of that answer. First, what exactly is being given away, because the term open weights is used loosely. Then the strategy that makes the giveaway rational, the peculiar economics that make it cheap, and the pushback that may yet change it.

02 What Open Weights Actually Are

Open weights are the publicly released learned parameters of a trained AI model - principally the weights and biases that encode everything the model learned during training. Publishing them allows anyone to download and run the model on their own hardware; whether they may modify it, fine-tune it for a specific task, or redistribute it depends entirely on the accompanying license. The practice is most common for large language models and other generative AI systems.

Three distinctions matter. Parameters are not the training code or the data; you receive the finished brain, not the recipe. A permissive license is not open source in the software sense; many open-weight releases restrict commercial use, derived model naming, or deployment scale. And having weights is not having the service - you must supply the computers to run them.

The distinction matters commercially because it defines what is being commoditized. If only access is free, the provider still owns the product. If the weights are free, the model itself becomes a commodity input, and value migrates to whoever applies it. That migration path is the entire strategic argument examined next.

03 The Chinese Open-Weights Stack

The ecosystem has names now. DeepSeek publishes frontier-grade models with unusually low reported training costs and detailed technical reports. Alibaba's Qwen family spans a wide size range and has become one of the most fine-tuned lineages in the world. Moonshot's Kimi models push toward very large parameter counts, and Zhipu's GLM series holds a significant share of domestic enterprise deployment. These are separate organizations with separate incentives, which strengthens the pattern.

Licenses vary by family and by size class, and the variation is deliberate: some releases permit unrestricted commercial use, others reserve restrictions for the largest tiers or for certain jurisdictions. Reading the license is not a legal formality but part of the engineering decision, because terms can differ from intuition about the word open.

What each lab ships, in practice, is a portfolio: total parameter counts for capability, active parameter counts for cheap inference, and small variants for local deployment. The chart below puts the flagships on one scale, with both dimensions labeled where disclosed.

04 The Strategy

The strategic logic has four moves. First, commoditize the model layer: if capable models are free, nobody can charge much for a model as such, and the profit pool shifts to applications, infrastructure, and integration - domains where Chinese companies are strong domestically.

Second, set standards. When developers around the world build on Qwen or DeepSeek architectures, tokenizers, and fine-tuning pipelines, those become the default vocabulary of applied AI, and defaults are a form of soft power that no export control reaches. Third, win developer mindshare especially outside the US: every university lab, startup, and government project that standardizes on an open-weight stack is a market closed to closed-API vendors by default rather than by decree.

Fourth, erode closed-API pricing power. If a free download reaches ninety percent of a closed model's capability, the closed vendor must price against zero, not against competitors. That margin compression applies to US labs regardless of where the open weights come from, which is why the giveaway is felt most sharply in San Francisco boardrooms.

05 The Economics

Training is a one-time capital expense; distribution is nearly free. The reported figures make the asymmetry concrete: DeepSeek-V3's compute was reported at roughly 5.6 million dollars, Llama 3.1 405B at around 60 million, and GPT-4 at over 100 million. Whatever these figures exclude, their spread shows the cost of capability is falling faster than most planning assumed - and a finished model, unlike a factory, can be copied at zero marginal cost.

Export controls restrict the compute side - advanced GPUs are physical goods subject to shipment controls - but weights flow over the internet as information, and the cost of a download does not scale with the value of the file. A trained model embodies compute that no longer needs the machine that made it.

Finally, giving away version N keeps optionality over version N plus one. A lab can publish this year's model as a standard-setting move and close a future one if revenue demands it; the license gate swings one way and can swing back. Interpretation, not fact: this reversibility is precisely what makes the strategy rational rather than ideological.

06 The Pushback

The first objection is misuse. Weights cannot be un-published, and safety training can be removed from an open model by fine-tuning - removable guardrails, in the blunt phrase. Once a file circulates, no lab retains the control it keeps over a hosted API, and the most capable freely downloadable model defines the floor for everyone, including the malicious.

Second, the Washington debate. Open weights from Chinese labs are discussed in the United States both as an economic threat to closed-lab business models and as a national-security question, with proposals ranging from licensing regimes for large model releases to nothing at all. The policy outcome is genuinely uncertain, which is itself a risk factor for any strategy built on the free flow of files.

Third, enterprise hesitancy. Procurement teams weigh license terms, indemnity, and provenance; a permissively licensed model from a little-known lab faces diligence friction that a commercial API does not. Licensing risk, not capability, is often the binding constraint on enterprise adoption of open weights, and vendors increasingly sell support and liability coverage on top of free models to close that gap.

07 What to Watch

Three indicators will tell you whether the strategy holds. First, license tightening: labs can ratchet terms on successive releases, restricting commercial use or jurisdictions while still calling the release open. Watch the diff between versions of the same family's license, not the marketing page.

Second, the temptation to go closed for revenue. If API revenue at any Chinese lab overtakes the strategic benefits of openness, or if domestic regulation repositions model releases as strategic assets, the current pattern inverts and the free frontier migrates elsewhere. The option to close is the strategy's silent cost.

Third, whether Western labs respond by opening up. A credible open-weight countermove from a US lab would reframe the whole contest as a race for mindshare rather than a one-directional giveaway, and would tell you the strategic reading in this article was right. If instead the free frontier simply persists for years, the answer to why is giving away was always: because it buys something better than money.