What does open-source AI mean?

The model's weights — the learned parameters that make it work — are published so anyone can download, run, and build on them. DeepSeek's R1 is released under the MIT licence and Alibaba's Qwen under Apache 2.0 (model cards).

Which Chinese AI models are open source?

DeepSeek and Alibaba's Qwen families are the best known. DeepSeek has released openly since January 2025, including V4-Flash in July 2026 and V4-Pro in August 2026 with million-token context. Their downloads number in the millions per month on Hugging Face (as-of 2026-08-29).

Why does OpenAI keep its models closed?

Closed weights protect the business — customers must pay the company to use the model, and competitors cannot copy it. GPT-OSS, OpenAI's 2025 open release, was an exception and performed well below its own closed models.

Is open-source AI free?

The weights are free to download, but running them requires hardware and skill. Apps like Plainly serve open-source models for $5 a month with no technical setup.

Can open-source AI models be taken away or shut down?

No. Once weights are released under a permissive licence, anyone can run the model on their own hardware forever. There is no company that can revoke access, change the price, or shut the model down.

Is GPT-OSS as good as DeepSeek?

No. GPT-OSS is okay, but it sits far below OpenAI's own closed models, while DeepSeek-class open weights are genuinely useful for everyday work. I have used GPT-OSS a couple of times and cannot recommend it.

Why does China open-source AI models while the US keeps them closed?

China's government pressures subsidised labs to release open weights, so anyone can run their models. The US keeps its frontier weights closed because access is the business. The two strategies compound differently — and the difference explains everything about AI pricing.

Why does China open-source its AI models?

My belief, stated as such: the Chinese government pressures its subsidised labs to open-source their weights. The result is that anyone in the world can download and run a DeepSeek or Qwen model. The US, with no equivalent pressure and a business model built on exclusivity, keeps its frontier weights closed. The two strategies are not just different — they compound differently, and the compounding is what explains why open models keep getting better faster than anyone predicted.

The two strategies, side by side

	China	US
Flagship families	DeepSeek, Qwen	GPT (closed), Claude, Gemini
Weights	Published openly	Kept closed
Who can run them	Anyone, anywhere	Paying customers, via the company
Business logic	State-backed; ecosystem strength over direct revenue	Private capital; access is the product
Ecosystem effect	Labs build on each other's releases; the whole field compounds	Techniques stay inside each lab
Risk to the user	Low — the model cannot be revoked	Ongoing — access depends on one company's prices and policies

Why open compounding matters

When a Chinese lab releases weights, it does two things at once, and the second is the underappreciated one.

First, it removes the company as a gatekeeper. You can run the model yourself, on your own hardware, forever. There is no vendor that can revoke access, raise the price, or change the terms. For any organisation that cares about not depending on a single American company, this alone is decisive.

Second — and this is why the open ecosystem moves so fast — competing labs study the released techniques and build on them. Each release raises the floor for every other lab. A training method published in one lab's release shows up in three competitors' models within months. Compare that to the closed world, where every technique stays inside the lab that discovered it and the field advances at the speed of its slowest internal roadmap.

The release cadence shows what compounding looks like in practice. DeepSeek has shipped openly since January 2025 — reasoning model, then a steady stream, including V4-Flash (31 July 2026) and V4-Pro (13 August 2026) with million-token context, under the MIT licence (Hugging Face model cards, as-of 2026-08-29). Alibaba's Qwen family ships under Apache 2.0. These are not curiosity releases; they are the engines of the cost collapse described in why Chinese AI models cost less.

The download numbers show where the gravity is: DeepSeek-R1 alone pulls over 2.7 million downloads a month on Hugging Face, and Qwen3-235B around 370,000 (as-of 2026-08-29). Millions of people and companies have decided the open weights are worth running.

The US counter-move — and why it is different

The West's one notable open release, OpenAI's GPT-OSS in August 2025, tells its own story. It is okay. It runs on modest hardware — the 120b variant fits on a single 80GB GPU — and OpenAI claims it performs around the level of its own o4-mini. But it sits far below OpenAI's own closed frontier, and I have used it a couple of times and cannot recommend it over the DeepSeek-class alternatives.

My belief, stated as such: GPT-OSS was a legal play rather than a strategy — released under pressure from the Musk lawsuit over the drift from OpenAI's charitable mission, to demonstrate openness without giving away anything commercially valuable. Whatever the motive, the pattern holds: the real US frontier stays closed, because closed is the product. Every dollar of that $852 billion valuation depends on access being something you pay for monthly.

The two "open" stories are not comparable, and it is worth naming that plainly when you see GPT-OSS cited as evidence that "everyone open-sources now".

What the split means for you

Open weights and closed weights are not competing products. They are competing futures: one where capability is infrastructure anyone can build on, one where capability is a subscription. For an everyday user, the practical effect today is simple — the models you use through a cheap app are open ones, and their quality improves on the open ecosystem's clock, which is fast.

The honest limit

Open does not automatically mean better. The closed US frontier still wins at the very top end, and for the hardest problems I still pay for it. Open wins on cost, transparency, and freedom from lock-in — which happens to be everything an everyday user actually needs. If you are deciding what to pay, is a $5 AI good enough works through what the open side covers and where it stops.