What is an open-weight model?
An open-weight model is a machine learning model whose trained parameters, the weights, are published for anyone to download and run on their own hardware. It's a claim about distribution, not about how the model was built: the training data, training code, and evaluation methodology can all stay private while the weights themselves are freely downloadable.
Weights are the output, not the recipe
A model's weights are the numbers learned during training, the parameters that turn an input into an output once the architecture is fixed. Downloading a model's weights gets you a working copy you can run yourself, fine-tune, or inspect. It doesn't get you the dataset that produced those weights, the training scripts, the compute budget, or the decisions made about data filtering and curation along the way. Those are separate artifacts, and most open-weight releases don't include them.
Why this differs from "open source"
Open source, applied to software, usually implies you can see and modify the thing that produces the artifact, not just the artifact itself. Applied loosely to models, "open source" often just means open-weight. That's a meaningfully weaker claim: you can run and fine-tune an open-weight model, but you generally can't reproduce it from scratch, audit exactly what it was trained on, or verify claims about data provenance. Some organizations do release training code and dataset details alongside weights, which is closer to the software sense of open source, but that's the exception rather than the rule among current releases.
Licensing is a separate axis from openness
A model being open-weight says nothing about what you're allowed to do with it. Licenses attached to open-weight releases range from permissive (Apache 2.0, MIT) to custom terms with commercial-use thresholds, redistribution limits, or acceptable-use restrictions tied to the model provider's own policies. Two open-weight models can sit on very different points of that spectrum, so the license text for the specific release matters more than the "open" label. For the practical question of when open-weight is the right choice over a closed API, see our comparison of open-weight vs closed-weight models.
What open-weight makes possible
Because you have the actual parameters, you can run the model on infrastructure you control, modify it through fine-tuning, quantize it to fit smaller hardware, or inspect its behavior directly rather than through an API's output alone. None of that is available with a closed-weight model, where the provider keeps the parameters on their own servers and exposes only an API. That difference is why self-hosting is only an option at all for open-weight models; see our glossary for related terms like quantization and inference.
Checking a release before you rely on it
Before treating any release as suitable for a given use, it's worth checking three things directly on the model's own page or repository: the license terms attached to that specific release, whether the publisher discloses anything about training data sources, and whether independent benchmark results exist beyond the publisher's own claims. Open-weight releases vary widely in how much of this is documented, and a model card that's silent on data provenance isn't unusual, but it's still worth noting before building something that depends on the answer.
Where GPUwerk fits
Running an open-weight model on your own hardware is exactly what a dedicated Spark is for. A GPUwerk Spark gives you 128GB of unified memory to load and serve an open-weight model directly, starting at $0.79/hour from EU-Central, with no vendor sitting between you and the weights. More detail is on the private LLM hosting page.