Open Source Series · What Open Source Actually Opens

Open Source Is a Business: What Each Vendor Is After

Training a flagship model burns an enormous amount of compute, and once the weights are out they can never be pulled back. So why do so many companies still open-source? The answer has nothing to do with charity. Understand what each of them is counting on and your selection decisions get a lot steadier.

Six vendors, six calculations
MetaOpen weights
It does not make money selling models; its core business is social advertising. Turning the model into a free commodity pulls the pricing floor out from under its rivals. Economists call this commoditizing your complement: the cheaper the complement, the more valuable your core business.
GooglePartially open
Small models go open source to win developers, while the flagship stays closed and monetizes through the API and cloud services. Two legs to stand on: it keeps a foot in the ecosystem without giving up pricing power over its strongest capability.
MistralGenuinely open
Apache 2.0 on the base models builds technical credibility, while revenue comes from enterprise customization and on-premise deployment contracts. It earns from services, not from API lock-in.
Alibaba · QwenFully open
Open source establishes a de facto standard: developers fine-tune and deploy, and eventually a large share of that compute demand comes back to the cloud. Free models, paid compute. That path only works if the ecosystem is large enough, which is why it has stayed aggressive on size coverage and release cadence.
DeepSeekMIT licensed
It uses MIT, about as permissive as licenses get, and explicitly allows distilling from its output. Very little marketing spend buys technical credibility worldwide: one release draws more attention than an equivalent budget spent on advertising.
OpenAIFully closed
It started as a nonprofit doing open research and moved to a closed commercial model. The strongest capability is the moat, and open-sourcing would fill the moat in, which conflicts directly with a subscription and API business.
The pattern: whether a company open-sources depends on where the model sits in its revenue structure. If the model is the product, it tends to stay closed. If the model is a customer-acquisition channel or a complement, it tends to be opened.

What does that pattern imply for you? Pick whichever revenue model is closest to how your own company makes money.

How does your company make money?
How to measure the real influence of an open-source ecosystem

Download counts are easy to inflate, and leaderboard positions shift the moment you change the test. What the industry weighs more heavily is the number of derivative models: how many developers actually took the weights, trained something new, and released it. That number cannot be faked, because behind every derivative model sits a real investment of compute.

200K
Derivative models
The first open-source model family in the world to reach this order of magnitude
1B
Cumulative downloads
About 1.1 million per day
~400
Models open-sourced to date
Since 2023, covering text, multimodal, code, and other directions
Figures are Hugging Face platform statistics for the Qwen family as of 2026-01-21, reported by Sina Finance, Yicai, East Money, and others. For comparison, the family's derivative model count stood at roughly 90,000 in February 2025 and passed 100,000 during 2025. Verified 2026-08-07.

What this set of numbers really shows is that developers vote with their feet. When they choose which base to keep training on, they weigh whether the license is clean, whether the size lineup is complete, and whether the community already has a working toolchain — none of which has much to do with marketing. So the derivative model count is essentially a long-term record of user choice.

Then someone filled in the moat

The six vendors above are counting on very different things, but on one point they were strikingly consistent: you can open-source the small models, but not the strongest one. Google released Gemma and kept Gemini. Meta gave Llama to everyone and kept its strongest capability inside its own products. OpenAI does not even publish architecture details. The reason never needed explaining: the flagship is the moat, and nobody fills in their own moat.

That rule held for more than three years. Then came this summer.

2026-08-03

For the first time, a flagship's weights were released

2.4 trilliontotal parameters
95 billionactive parameters
Apache 2.0open weights, no added conditions

Qwen3.8-Max was open-sourced. It is the first Max-tier flagship model with open weights, and there is no prior precedent for it.

What that means is that anyone can download a model at the current frontier scale, take it apart to see how it was built, modify it, and keep training on it, without asking anyone for permission. The capability gap between the open camp and the closed camp, the one everyone had assumed was structural, does not hold anymore.

What makes this interesting is that the move does not break the business logic in the table above. It pushes that logic to its end point. Free models, paid compute only works if the ecosystem is large enough, and if that is the path you are on, releasing the strongest model too is the most aggressive move available and also the most rational one.
Source: Qwen official blog, Qwen3.8-Max release notes, 2026-08-03. 2.4 trillion total parameters, 95 billion active, released under Apache 2.0. Verified 2026-08-07.

Now a bucket of cold water: you almost certainly cannot run this. Weight files for 2.4 trillion parameters run to several terabytes, and the calculator at the end of this chapter will tell you it won't run. So its practical significance is not whether you can open it on your own machine. It lies in three other things.

1

Researchers can dissect a model at this scale for the first time

Models this large could previously only be guessed at from behind an API. Now the weights can be read directly, and for a lot of questions about how large models actually work, the conditions for verifying an answer exist for the first time.

2

Real on-premise deployment now has a much higher ceiling

Organizations whose data cannot leave the internal network used to be choosing among small models only. Now, as long as they can afford the hardware, their capability ceiling is in the same league as using the strongest API.

3

It can be a teacher

This is the one that matters most to you. A top-tier model with open weights means the teacher for distillation no longer has to be somebody else's API. The second half of this chapter covers distillation; come back to this then and weigh what it is worth.

Back to the three-tier table from the previous section. Qwen3.8-Max lands in tier one: standard open-source license, downloadable weights, commercial use allowed, training allowed. Not a word of the criteria has changed. What changed is that this time, the thing landing in tier one is a flagship.
For the people making the selection

Use this section together with the previous one:

Start with the license to confirm whether you can use it and what conditions apply.
Then look at the business logic to judge whether the vendor will keep open-sourcing. Companies that earn directly from the model can tighten their open-source policy at any time; companies for which open source is a customer-acquisition channel tend to be more stable.
Finally look at ecosystem depth. Plenty of derivative models and a complete set of community tools means you can find answers when you run into trouble, and that someone has already hit the bug ahead of you.
Concentration has a price. When two hundred thousand derivative models all grow out of a handful of bases, they inherit the same biases, the same phrasing habits, and even the same mistakes together. That is the problem the second half of this chapter deals with specifically.