First, I want to take you back a few years. It was March 2023. ChatGPT was everywhere. Then a rumour and a torrent file started to move around social media. People who downloaded this file found something interesting inside. It was the weights of LLaMA, Meta’s large language model. Meta gave these weights only to researchers who applied for them. But now the model was public.
Think about it. This looks like a big crisis for a company. But Meta turned the crisis into a strategy. In July 2023, Llama 2 came out as an open weight model. Now it was open to everyone. Zuckerberg wrote articles and gave many interviews. He said that open source makes better and safer models. For two years after Llama 2, Meta was on the other side of the closed models. You could download the model. You could change it. You could even run it on your own server.
This continued until April 2026.
The logic of the gift
Meta’s open weight strategy was never charity. If you think a little, the goal becomes clear. Meta wanted to turn the most valuable asset of its rivals into a cheap product. It wanted to build an ecosystem around its own model. And it wanted to attract good engineers. To be fair, all of this worked.
But there was one big problem. A model that everybody downloads to their own computer does not produce data for Meta. When Llama runs on your computer, Meta knows nothing about that use. It does not see the user’s prompt. It does not see the model’s answer. It does not know where the model failed, and it does not know how the user fixed it. Meta did a lot of work and got almost zero feedback.
In 2023, nobody thought about this cost. Models used public internet sources for training. At that time, there was a lot of usable data in those sources.
What is not on the web
Epoch AI made a calculation. They say that public, human-made, quality-checked data, in other words organic data, is about 300 trillion tokens. At this speed, language models will use all of this data between 2026 and 2032. This looks like a wide range, but we also continue to produce organic data.
However, we should look at the type of this data, not the amount. The public web is really an archive of results. Blog posts, product documents, answers on StackOverflow. All of them are finished work. None of them tell us how the work was done. The three attempts that took hours, the bugs fixed at two in the morning, the discussions about them. We never had access to these. But this is exactly what agents need to learn.
Mario Zechner is the developer of an open source harness tool called Pi. He has an observation. He says agents made a big jump between April and October 2025 because Claude Code saved all sessions by default and used them for reinforcement learning. Of course, this claim is not confirmed. It is the observation of a developer who works in this area. But nobody in the industry said the opposite either.
Meta also has a great treasure, but it is the wrong type. Facebook, Instagram and WhatsApp produce a huge amount of data. But they produce almost no coding data, and agents need coding data. Also, Meta’s open weight strategy closed the only channel that could fill this gap. Meta did this with its own hands.
The break
Let me give you a short timeline. Llama 4 came out in April 2025 and did not meet expectations. In June 2025, Meta paid $14.3 billion for 49 percent of Scale AI and hired Alexandr Wang. Superintelligence Labs was created. In April 2026, Muse Spark came out. It was Meta’s first closed weight frontier model.
Meta did not even explain the move to closed weights. But Meta also had to do something about the data that this new model needed. To close the gap, it installed monitoring software on the computers of its employees in the United States. The program was called Model Capability Initiative. It collected mouse movements, clicks, keyboard activity and screenshots from time to time. The goal was to teach the AI how people do their work. On 22 June 2026, the program was stopped. A configuration mistake made the sensitive employee data open to all Meta staff.
When this road closed, only one option was left. Meta had to buy the data it needed.
Two different prices
Two versions of Muse Spark came out in August and September. But the interesting part is the price. These models have two different prices.
In the standard version, one million input tokens cost $1.25 and one million output tokens cost $4.25. Meta says that your prompts and the model’s answers will not be used to train this model or future models.
The other version is called Contributor. Here one million input tokens cost $0.10 and one million output tokens cost $0.20. In return, Meta gets the right to train its future models with your data. That is 92 percent cheaper for input and 95 percent cheaper for output. It is very cheap, even much cheaper than the Chinese models.
One more thing is important. The technical features of the two models are exactly the same. The context window is the same. Tool calling is the same. The training data is the same. The only difference is the price. And that difference is the value of your usage data.
The Glimmer test
In fact, Meta did not leave open weights completely. On 10 August, a Muse variant called Glimmer was introduced. Muse Glimmer has 30 billion parameters. It uses the Apache 2.0 licence and it can run on a single GPU. Its licence is more generous than the Llama licence. So Meta came back to open weights with a more open licence than ever before.
But Glimmer is not the flagship model. It is distilled from Muse Spark. It is one class lower, but it can run on local hardware. In other words, it is the model that teaches Meta nothing. The model that teaches Meta is closed. The model that teaches Meta nothing is Apache 2.0.
Wang also promised an open weight version of Muse Spark. He repeated this promise three times: on 10 August, on 20 August and at the launch on 2 September. Until now, nothing has happened.
I think this new strategy will be more successful than the first one. This mixed approach may also inspire other companies in the future.
Direction
Meta never said that it left open weights because it could not get data. This is my own conclusion from the sources. But everything points to this. The company that shared its model with open weights closed it after three years. Then it tried to monitor its employees. It failed. And then it started to buy data from outside and pay for it.
With open weight models, when you download the weights, everything is yours. Meta or anyone else does not know what you do with it and learns nothing from it.
In Meta’s new pricing, the part of the bill that you do not pay is a gift. But this time, you are the one who gives the gift.
Sources
TechCrunch, “Meta is paying to peek at how you use their latest AI model,” 3 September 2026 - techcrunch.com
Meta Model API, Pricing & Rate Limits - dev.meta.ai
Meta AI Research, “Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device,” 10 August 2026 - research.meta.ai
Reuters, “Meta to pause internal mouse-tracking tech while examining data security issues,” 22 June 2026 - reuters.com
The Register, “Zuck rekindles open weights Llama drama with Muse Glimmer,” 10 August 2026 - theregister.com
Epoch AI, “Will we run out of data? Limits of LLM scaling based on human-generated data,” June 2024 - epoch.ai