Plain-English explanation
Some AI models are only accessible through an API: you send a prompt to a company's server, and you get a response back, but you never receive or handle the model file itself. Other models are released with their weights published, typically on a hosting platform, for anyone to download and load into their own software on their own hardware. Those are open-weight models. The distinction is entirely about access to the weights themselves, not about how the model was originally built or trained.
Adult-AI use
This distinction is one of the most practically important in the whole adult AI space, because it's what determines whether local generation is even possible for a given model. Open-weight image models and language models (often published on hubs like Hugging Face or community sites like Civitai) are what get downloaded as checkpoints, run through local generation software, customized with LoRAs or full fine-tuning, and used without being subject to a hosted platform's content policy or filters, which is a large part of why the local, uncensored end of adult AI generation and companion apps depends on open-weight models existing at all. Hosted-only, closed models can't be run locally, customized this way, or freely redistributed, regardless of how capable they are.
How it differs from "open source"
Open-weight is often used loosely as if it meant "open source," but they're not the same thing. Open source, applied strictly to software, usually implies the full recipe is available (training code, and often the training data) allowing someone to reproduce the model from scratch, not just run it. Most open-weight model releases publish only the finished weights (and often a permissive-but-not-fully-open license) without the underlying training data or complete training pipeline, so the model can be run, quantized, and fine-tuned, but not necessarily reproduced or fully audited from raw materials.
Common uses
- Running an image or language model entirely on local hardware without depending on a hosted service.
- Applying custom fine-tuning, LoRAs, or other adaptations that hosted-only models don't allow.
- Converting a model to a different format or precision, such as quantizing it or converting it to GGUF, to fit consumer hardware.
- Building products or apps on top of the model without paying per-request API fees to a model provider.
Limitations and misconceptions
- Open-weight isn't automatically unrestricted. Most open-weight releases still come with a license governing commercial use, redistribution, or certain applications, even though the weights themselves are downloadable.
- Running it still takes real hardware. Access to the weights doesn't remove the need for a capable GPU (or patience, plus quantization) to run larger models locally at reasonable speed.
- Open-weight doesn't mean untrained-on-anything-sensitive. The training data and process behind an open-weight model is often not disclosed even when the resulting weights are.