What Open-Weight AI Models Are, and Why the Label Matters

When a company releases an AI model as open-weight, it makes the trained numbers that define the model available for anyone to download. Those numbers, called weights, are the parameters a system learned during training, and together they determine how the model turns an input into an output. With the weights in hand, developers can run the model on their own hardware, inspect how it behaves and fine-tune it for specific tasks, without paying per request to the company that built it.

The term is narrower than it sounds. Closed models, such as those offered through a subscription or an API from a major lab, keep their weights private and let you send in questions and receive answers. Open-weight models hand over the artifact itself. But the release often stops short of full open source. Companies may withhold the training data, the exact training code or the recipe used to build the model. Many also attach custom licenses that limit commercial use above a certain user count or bar certain applications. Researchers have pushed for more precise labels for this reason, since "open" can suggest that everything is public when it is not.

Notable examples include Meta's Llama family, Google's Gemma models, Mistral's releases from France, Alibaba's Qwen series and DeepSeek's models from China. OpenAI also released open-weight models in 2025, a notable shift for a company better known for closed systems. Availability changes quickly, so any list like this will look different over time.

The appeal is practical. A hospital or law firm can run a model on its own servers so sensitive records never leave the building. A startup can avoid per-token fees that climb at high volume. Researchers can study the internals, test for failure modes and reproduce results. Hobbyists can run smaller versions on a gaming PC. Fine-tuning lets a company teach a model its own terminology or house style at a fraction of the cost of training from scratch.

The tradeoffs are real. Large models need substantial memory and GPU capacity, and running them well takes engineering time for hosting, updates and monitoring. Safety is also harder to guarantee. Anyone with the weights can fine-tune away safeguards the original developer added, and once weights are public they cannot be recalled. Teams weighing these models usually balance those risks against the benefits of control and transparency.

If you are evaluating a model, check three things: the license terms, whether the training data and methods are documented, and whether the weights can actually be downloaded without a waiting list or a gated form. The open-weight label is a useful signal, but the details decide how open a model really is.