Open Weight LLMs are NOT Open Source Software

Meta released a new open weight LLM today, called Muse Glimmer. They also released a paper that says, basically, that data wants to be free. Fundamentally, the paper says that a wider dissemination of software is good instead of under the control of just a few people like, say, his arch enemies, Anthropic and OpenAI.

The New York Times reported on this and that reporting was repeated on PBS tonight and they sort of got some things right and some things wrong.

One thing they got right is that Meta’s pitch is self serving. Sort of. More later.

The most important distinction is that open weight and open source are two very different animals. Importantly, what Meta released is not software that the average person can use. What they released are the weights that the Muse Glimmer model is based on. With Glimmer, you will need to have a powerful enough computer, an LLM framework like Ollama (which is Open Source) and the skill to know what to do with it. For most people this is out of their reach.

So, having the weights, basically, the parameters, allows developers to develop software around it, but it is, in and of itself, not software as the average person considers it. So part of what Mark Zuckerberg said is right. IF developers CHOOSE to write software around Mark’s parameter dump, it MAY make AI more widely available. ASSUMING users think the software that random developers may or may not write is useful. And that software may or may not be free.

One of the premises that open source advocates point to is that if the software is available for anyone to look at, bugs will be found more quickly. As a philosophical concept this is true. But the most popular open source software in the world is Linux, which has been around for a couple of decades and is widely studied. We continue to find bugs in it all the time. Recently, due to that same AI, a number of very critical bugs have been discovered. That means, and this is important, that open source does not mean bug free. Not even remotely.

Also, in the case of what Meta released, it is not software at all, so there is nothing to look at to find bugs.

Worse yet, the training data was not released so we have no way to know how biased it might, possibly, be. Hence, NOT OPEN SOURCE.

So why did Zuck do what he did? Meta has spent billions on their AI. Almost no one uses it other than them. OpenAI and Anthropic, among others, are cleaning his proverbial clock in this arena. Mark doesn’t like losing. If he can do something to hurt his competitors – and giving away something that his competitors are selling is a good way to do that, he has shown that he is more than willing to do that.

Another point the Times made is that Anthropic and OpenAI say these new models are too powerful. Bad actors will use them so they, the guardians of the (AI) galaxy, are here to protect us. While this is partially true, of more interest to them is that the guardians want to be in control and closed software allows them to do that and make money at the same time. If Mark gives away the weights, he makes zero money. Which, for now, might be okay since he makes bazillions from selling ads on his platforms. OpenAI and Anthropic don’t have those revenue streams so giving away their software is “problematic” for them.

Next, the Times repeated OpenAI and Anthropic talking points (not taking their position but just communicating it) that bad actors will use the software in evil ways. Absolutely true. Cars are used by bank robbers to get away, but we don’t ban cars. These two companies locking down access to their software won’t stop that. There are many open-weight AI models already available. One, from China, called Kimi K3 is almost as powerful as OpenAI and Anthropic frontier models and closing the gap fast. If the goal is to make hackers use Chinese software to hack our businesses instead of using American software to hack us, that appears to me to be a difference without a distinction.

While I could go on for a long time, I will get off my soap box with just one more item. Meta’s Glimmer has just shy of 30 billion “parameters”, a measure of how powerful the AI is. On the other hand, Mythos from Anthropic has between 5 and 10 trillion parameters with about 800 billion to 1.2 trillion active for any one query. While OpenAI has not released numbers, estimates are that the GPT-5 family has between 10 and more than 50 trillion parameters with 2 to 5 trillion active for any one query. The free Chinese open weight model Kimi K3 has 2.8 trillion parameters. Parameters are a very rough correlation to power or usefulness. VERY. ROUGH. That means that Mark’s 30 billion parameter model is designed to run on a computer that a business can afford to buy, but is way less powerful than either of his big US competitors.

This doesn’t mean that Glimmer is useless. But it is important to understand what you are getting. You can find some reporting on the Times reporting at Gary Marcus’ blog. Glimmer is not going to kill Anthropic or OpenAI any time soon.

by