What Is an Open-Weight Model?

An open-weight model is an AI model whose trained parameters, the weights, are published so anyone can download and run it. You can host it on your own servers, keep every prompt inside your network and tune it on your own data. What you usually do not get is the training data or the full training code, which is why “open-weight” and “open-source” are not the same thing.

Open-Weight vs Open-Source

The Open Source Initiative’s Open Source AI Definition says an AI system is open source only when users can use, study, modify and share it, and when three things are available: detailed information about the training data, the complete code used to train and run it, and the parameters. Most popular open-weight models publish the parameters and inference code only, under a licence written by the vendor.

Closed (API only)Open-weightOpen-source AI (OSI)
Weights downloadableNoYesYes
Training codeNoRarelyYes
Training data informationNoPartial at bestYes
LicenceVendor terms of serviceVendor licence, often with conditionsOSI-approved terms
Where it runsVendor cloudAnywhere you chooseAnywhere you choose

Read the licence before building on a model. Some open-weight licences limit commercial use above a user threshold, forbid certain uses or require attribution. Others use a standard permissive licence.

Why Teams Choose Open-Weight Models

  • Data control. Prompts and documents never leave your infrastructure, which is often the simplest route to meeting data residency and confidentiality requirements.
  • Stability. A downloaded model does not change or disappear when a vendor updates its API.
  • Customisation. You can fine-tune, quantize or distill it freely, within the licence.
  • Cost at volume. At high, steady traffic, owning GPU capacity can cost less than paying per token.
  • Offline and edge use. A small open-weight model can run on a device with no network.

Costs and Risks

  • You own operations. GPU capacity, scaling, patching, monitoring and security are your team’s job.
  • Quality gap on hard tasks. On the hardest reasoning tasks the best hosted models are often ahead. Measure on your own tasks before assuming either way.
  • Safety is your responsibility. Vendor-side filters do not come with the weights. You add guardrails yourself.
  • Supply chain. Weights downloaded from public hubs are files from a third party. Verify the source and checksums, and load them in formats that cannot run arbitrary code.
  • Regulation still applies. Self-hosting does not exempt a system from the EU AI Act or from data protection law.

The Policy Debate

Open weights are also a policy question, because once released they cannot be recalled. In its July 2024 report on dual-use foundation models with widely available weights, the US National Telecommunications and Information Administration (NTIA) concluded that current evidence did not justify restricting them, and recommended that the government actively monitor the risks and keep the option to act later.

An open-weight model can be a large language model or a small language model; both are foundation models. Many companies use both open-weight and hosted models in one product, with model routing deciding which one handles each request.

In Practice

An open-weight model is chosen when data control or cost at volume justifies running it, and the choice is proved on an evaluation set of real inputs against a hosted model. The comparison counts the operating load too: GPU capacity, the serving stack, security patches, and who upgrades the model when a better version appears. Each model’s license is read before it is shortlisted, since some open-weight licenses restrict commercial use or the number of users.

We deploy self-hosted models in private LLM and data residency projects, run them under LLMOps services and put small ones on devices through edge AI development.

Back to Glossary

Let's talk about your project

We'll help you harness AI and ML technologies to create innovative software solutions

File requirements: pdf, doc, docx, rtf, ppt, pptx — max 15 MB

No file selected