Open-Weight Models

Open-weight models are AI models released with their weights available for download. The weights are the billions of numbers learned during training that make a model work. With access to them, an organization can run the model on its own infrastructure, fine-tune it on its own data, quantize it to run on smaller hardware or host it through any provider that supports it. Many capable language models from technology companies and research labs are released this way, in sizes ranging from small models that run on a laptop to very large models that need a cluster of GPUs.

Open-weight models have reshaped the economics of AI. They give organizations an alternative to paying per token for proprietary models, they create competition among hosting providers and they allow deep customization. They are not free to use, however: someone still has to pay for the compute to run them, and the cost of hosting, operating and maintaining them can be substantial. For IT financial management, open-weight models expand the range of options for every AI workload and are central to decisions about self-hosting, fine-tuning and vendor dependence.

Definition of Open-Weight Models

An open-weight model is a trained AI model whose parameters are made publicly available for download and use. The developer may also publish code for running it, documentation and evaluation results. What is published varies from one release to another, so it is worth checking what each release includes.

The term is distinct from open source. Open-source AI, in the strict sense, would also include the training data and full training process, which most open-weight releases do not provide. Many open-weight models also carry licenses with conditions that differ from traditional open-source licenses.

License Terms

Open-weight licenses range from very permissive, allowing almost any commercial use and modification, to restrictive, limiting use by large companies, prohibiting certain applications or requiring attribution.

Some licenses restrict using a model's outputs to train other models, or require that derived models carry the same terms.

Legal review of the license is an essential step before using an open-weight model in a commercial product or at scale. The terms can also change between versions of the same model family, so each new version should be checked.

Ways to Use Open-Weight Models

Organizations can run open-weight models on their own hardware or cloud GPUs, as a self-hosted LLM, giving them full control.

They can also use them through inference providers that host popular open-weight models and charge per token, often at lower prices than proprietary models of similar capability.

Open-weight models are also the usual starting point for fine-tuning and model distillation, since their weights can be modified directly.

Importance of Open-Weight Models in IT Financial Management

Open-weight models give organizations choices beyond the price list of a single provider. The same model can often be obtained from several hosting providers or run in-house, which creates price competition.

For high-volume, steady workloads, hosting an open-weight model can cost much less than paying per token for a proprietary model, provided the infrastructure is well used.

They also reduce dependence on any one provider. If prices rise or terms change, an organization using an open-weight model can move it to another host without changing the model itself. That portability strengthens its position when negotiating with providers.

Cost Considerations

The cost of an open-weight model is the cost of running it: GPUs, serving software, operations staff and the effort of evaluating and updating models.

At low volumes, these costs can exceed what a proprietary API would charge, especially if dedicated GPUs sit idle much of the time. Using an inference provider that hosts the same model avoids this problem while keeping many of the benefits.

Comparing options requires calculating cost per task under each approach at realistic volumes, including quality differences between the models.

Control and Compliance

Running an open-weight model in-house keeps data within the organization's environment, which can be essential for regulated or confidential information.

It also gives control over when models change. Proprietary models may be updated or retired by providers, while an open-weight model can be kept exactly as it is for as long as needed.

These benefits have financial value, such as avoiding compliance risk and the cost of revalidating applications after provider changes.