An open-weight model is an AI model, most often a large language model (LLM), whose trained weights (the numbers the model learned during training) are published so anyone permitted by the license can download, run and adapt it on their own hardware or chosen cloud. By contrast, a closed model is typically available only as a service, through the developer’s own app or API or through partners it licenses. Open-weight does not automatically mean open source: the license, and how much of the training data and code is shared, varies from model to model.
At a glance
- The model’s weights can be downloaded and run outside the developer’s service, on your own servers, a private cloud or a hosting provider.
- Licenses vary widely: some are permissive, some restrict commercial use, scale or specific uses.
- Training code and detailed information about the training data are often not released, which is a key difference from the OSI’s definition of open-source AI.
- Running an open-weight model shifts hosting, security, updates and cost management to you or your provider.
- They are a common basis for private AI deployments.
What problem it solves
Closed AI models are easy to start with but come with trade-offs: your prompts and data go to the provider, you depend on their pricing and roadmap, and you cannot inspect or change the model. Open-weight models give organizations another option. They can run a capable model inside their own environment, keep data from leaving it, tune the model on their own data and switch hosting providers without changing the model.
For regulated industries, or for workloads with strict data residency or confidentiality needs, that control can make AI use possible where sending data to a public service is hard to approve.
How it works
The developer trains the model and publishes its weights, usually with documentation and a license. A user downloads the weights and runs them with an inference engine on suitable hardware, typically GPUs, either on premises, on bare metal or in the cloud, or through GPU as a service (GPUaaS). Smaller models can run on a single server or even a laptop; larger ones need significant GPU capacity for AI inference.
Organizations can use the model as-is, fine-tune it on their own data, or combine it with their documents using retrieval techniques. Many cloud and AI platforms also host popular open-weight models as a managed service, so you can use them through an API without running the infrastructure yourself.
The license sets the rules. Terms commonly address commercial use, user or revenue thresholds, attribution, acceptable-use restrictions and whether derived models must carry the same license. Terms differ from model to model and can change between versions.
When it matters for buyers
- When data cannot leave your control. Self-hosting keeps prompts and outputs in your environment, subject to how you secure it.
- When usage is high and steady. At volume, running your own model can cost less than per-token API pricing, but a fair comparison has to count hardware, staff and operations.
- When you need customization. Fine-tuning is often easier with open weights.
- When choosing between AI providers. Hosted open-weight models can reduce lock-in to a single model developer.
- For compliance. License terms, model provenance and acceptable-use clauses should go through your AI governance review and counsel.
Questions to ask vendors
- Which open-weight models do you host or support, and which versions?
- Who is responsible for reviewing and complying with each model’s license?
- Where will the model run, and does any data leave our environment or region?
- What hardware or GPU capacity is needed for our expected usage, and what does it cost?
- How are model updates, security patches and safety controls handled?
- What performance and accuracy results can you show on workloads like ours, compared with closed models?
How it differs from open-source AI
Open-weight describes access: the trained weights are published and can be downloaded. Open-source AI, as the Open Source Initiative (OSI) defines it, sets a higher bar. Besides the freedom to use, study, modify and share the model for any purpose, the OSI’s definition calls for three components: the model parameters, the complete training and inference code, and information about the training data detailed enough for a skilled person to build a substantially equivalent system. It does not necessarily require publishing the raw training data or allowing exact reproduction. Many models described as “open” release weights but withhold training code or detailed training-data information, or add license restrictions on use, scale or commercial purpose, so they are open-weight but may not meet that definition. Whether a particular model counts as open source is often debated. For buyers, the practical step is the same either way: read the license and confirm what it allows for your intended use.