An open-weight AI model is one where the company that trained it releases the finished model file for anyone to download, run on their own hardware, and often modify, instead of keeping it locked behind a paid API. This differs from a closed model, which you can only access through the company's own service, and from true open source, which would also require releasing the training data and code.
In this article
What "weights" actually means
Every neural network, including the large language models behind tools like chatbots and image generators, is built from millions or billions of numerical parameters called weights. During training, a model adjusts these weights repeatedly, based on huge amounts of example data, until it gets reliably better at predicting the next word in a sentence or the next pixel in an image. Once training finishes, those weights are essentially the model's memory, frozen into a file that can be copied, shared, or run on different hardware.
When a company releases a model's weights, they're handing over that finished, trained file. Anyone with the right hardware can load it and run it themselves, without sending data to the original company's servers. That's the entire mechanical difference between an open-weight model and a closed one: whether you can download the finished brain, or whether you can only talk to it through someone else's front door.
Open-weight versus closed versus open source
These three terms get used interchangeably in casual conversation, but they describe genuinely different things. A closed model, like the ones powering most major commercial chatbots, is only accessible through an API or app the company controls. You send a request, their servers process it, and you get a response back. You never touch the actual model file, and the company can change, restrict, or shut down access at any time.
An open-weight model gives you the trained file itself. You can run it on your own computer or cloud server, inspect how it responds to different inputs, and fine-tune it on your own data. What you typically don't get is the original training data or a full account of exactly how the model was built, which is why researchers avoid calling this "open source" in the strict software sense. True open source AI, where the training data, code, and weights are all public, is rarer, partly because training data often includes licensed or scraped content companies are reluctant to fully disclose.
Why companies release models for free
Training a large, capable AI model can cost tens of millions of dollars in computing power alone, so giving away the result seems like an odd business decision at first glance. Meta's release of its Llama models is probably the most-cited example, and the company's own stated reasoning has centered on building a broad developer ecosystem around its models rather than Google's or OpenAI's, similar to how giving away the Android operating system helped Google dominate mobile long before it made real money from it directly.
Other companies release smaller or older models openly while keeping their most capable, newest models closed and paid. This lets them capture research attention and developer goodwill from the open release while still monetizing their frontier work. There's also a talent angle: publishing open research and models is one of the ways AI labs compete for the relatively small pool of top researchers, many of whom prefer working somewhere that publishes and contributes back to the field.
Handing over the weights means handing over the finished brain, not just a door you can knock on.
Who actually benefits from open weights
Startups and independent developers get the most obvious benefit: the ability to build a product on top of a genuinely capable model without paying per-request API fees to a large company, and without sending user data to that company's servers at all. A healthcare startup handling sensitive patient information, for example, might specifically need a model that runs entirely on infrastructure they control, which a closed API can't offer no matter how good its privacy policy sounds on paper.
Researchers benefit too. Academic labs studying how these models work, where their biases come from, or how to make them more efficient generally can't do that kind of deep investigation on a closed model they can only send requests to. Open weights let a university lab actually look inside. Countries and companies concerned about depending entirely on a small number of U.S. tech giants for critical AI infrastructure have also pushed open-weight adoption for reasons closer to national or organizational independence than pure cost savings.
The licensing details that trip people up
"Open" doesn't mean "no restrictions," and this is where a lot of confusion happens. Meta's Llama models, for example, are free to use commercially for most companies, but include a clause requiring a separate license agreement from Meta if your product has more than 700 million monthly active users, a threshold aimed squarely at preventing direct rivals like Google or Amazon from building on top of Llama without a deal. Other open-weight releases carry research-only licenses that technically prohibit any commercial deployment at all, something developers occasionally miss until a legal team catches it later.
Some licenses also restrict specific use cases, like military applications or certain categories of content generation, regardless of whether the use is otherwise commercial. Reading the actual license text, not just the marketing language calling something "open," matters more than most developers realize before they build a product roadmap around a specific model.
The real risks worth understanding
Open-weight models can be fine-tuned by anyone with the right hardware and technical skill, which includes the ability to remove safety training a company originally built in. A closed model's safety guardrails live behind an API the company can patch and monitor; an open-weight model's guardrails, once downloaded, exist entirely on hardware the original company no longer controls. This has become a genuine point of tension in AI policy debates, with some researchers arguing certain categories of highly capable models shouldn't be released openly at all until safety research catches up.
At the same time, open access has produced real safety benefits. Independent researchers have used open-weight models to identify bias patterns, jailbreak vulnerabilities, and failure modes that companies with closed models had no external pressure to disclose. Our Technology coverage and Business section both track how this tension between openness and control is shaping product decisions across the industry, and it's likely to keep shifting as more governments weigh in with formal AI regulation.
How to evaluate a model before you build on it
If you're a developer deciding whether to build on an open-weight model, a few practical checks matter more than the headline benchmark scores. Check the license first, in full, not just the summary blog post announcing the release, since the summary rarely covers usage thresholds or restricted-use clauses in detail. Check how recently the model was updated and whether the releasing organization has a track record of maintaining older releases, since an abandoned open-weight model still works, but you won't get security patches or bug fixes if something is later discovered.
It's also worth checking the model card, a standardized documentation format most major labs now publish alongside a release, which typically discloses known limitations, the general categories of data used in training, and intended versus unintended use cases. A model card that's thin or missing entirely is itself useful information, since it suggests the releasing organization hasn't invested much in responsible deployment documentation, which can be a signal worth weighing alongside the model's raw capability.
Frequently Asked Questions
Is an open-weight model the same as open source?
Not exactly. Open source traditionally means the full source code and training process are public and freely modifiable. Most open-weight AI releases share the trained model file itself, letting you run and fine-tune it, without publishing the training data or the full code used to build it, which is why researchers use the more precise term "open-weight" instead of "open source."
Can I legally use an open-weight model to build a commercial product?
It depends entirely on the license attached to that specific model. Some, like Meta's Llama models under certain conditions, allow commercial use with a few restrictions, such as a user-count threshold that requires a separate agreement. Others are released under more restrictive research-only licenses. Always check the specific license before building a product on top of any open-weight model.
Are open-weight models less safe than closed models?
Safety tradeoffs run in both directions. A closed model's safety measures live entirely behind an API a company controls and can update. An open-weight model's safety training can potentially be removed by anyone with the technical skill to fine-tune it, but the open weights also let independent researchers audit the model for bias and vulnerabilities that a closed company might never disclose.
Why would a company give away a model that cost millions of dollars to train?
Several business reasons drive this. Some companies use open releases to build developer goodwill and adoption, hoping to sell related infrastructure or premium versions later. Others release older or smaller models openly while keeping their most capable models closed, using the open release partly as a recruiting and reputation tool in a competitive research field.
The Takeaway
Open-weight AI models hand you the trained model file itself, letting you run and modify it on your own hardware, which is a meaningfully different thing from either a closed API or true open source software. Companies release them for a mix of ecosystem-building, research reputation, and competitive strategy, not pure generosity, and the specific license attached to any given model matters more than the word "open" suggests on its own. Before building anything real on top of an open-weight model, read the actual license terms rather than assuming "open" means "no restrictions."