Where can I get access to open-source AI models?

Summary

• Open model weights are distributed through model hubs such as Hugging Face, through developers' own repositories, and through hosted services. You can download them and run them yourself, or call a hosted copy through an API.
• Licenses differ by model. OpenAI's gpt-oss is released under Apache 2.0, while the Llama 4 Community License requires "Built with Llama" attribution on distribution and a separate license from Meta above a user threshold. Read the license before you build.
• On Databricks, Foundation Model APIs give you open models hosted by Databricks behind a serving endpoint, with pay-per-token, provisioned throughput and batch options.
• You can also download a base model from Hugging Face or the Databricks Marketplace, register it in Unity Catalog, and serve it with provisioned throughput if its architecture is supported.
• Unity Gateway routes requests to models and applies Unity Catalog permissions, rate limits and usage monitoring. It has been generally available since August 4, 2026.

Where can I get access to open-source AI models?

Download the weights from a model hub such as Hugging Face or from the developer's own repository, then run them with a local runtime such as Ollama, llama.cpp or vLLM. Or call a hosted copy through an API. On Databricks, Foundation Model APIs host open models for you, and you can register models you download in Unity Catalog and serve them. Check each model's license first, because "open" covers very different terms.

What counts as an open-source model?

The Open Source Initiative's Open Source AI Definition 1.0 calls for terms that grant the freedoms to use, study, modify and share an AI system. It requires the model parameters, the complete source code, and detailed information about the training data. Releases vary in what they publish. OpenAI describes gpt-oss as open-weight models under the Apache 2.0 license, and Meta ships Llama under its own community license. Whether a given release meets the definition depends on what its authors publish and on its license.

Where can you find and download open models?

Source	What it offers	Worth knowing
Hugging Face Hub	Hosts model checkpoints for storage, discovery and sharing. Download with the huggingface_hub library, Transformers or other integrated libraries	Some models are gated: you share your username and email with the authors to get the files
Developer repositories	The gpt-oss GitHub repository describes the models, links to the weights on Hugging Face, and shows how to run them with vLLM, Ollama and Transformers	Read the license that ships with each release
Local runtimes	Ollama has a model library and runs on macOS, Windows, Linux and Docker. llama.cpp can download a model directly from Hugging Face. vLLM is a library for LLM inference and serving	You supply the compute
Databricks	Foundation Model APIs host open models. The Databricks Marketplace offers base models, and the system.ai schema in Catalog Explorer lists ready-to-use models	Governed by Unity Catalog

How do you run an open model on Databricks?

Databricks describes Foundation Model APIs as a way to access and query open models hosted by Databricks from a serving endpoint. There are three modes. Pay-per-token is the easiest way to start. Provisioned throughput is recommended for production workloads. AI Functions optimized models are recommended for batch inference. Most of these APIs are compatible with the OpenAI client. The supported-models list includes endpoints for Meta Llama, OpenAI gpt-oss and Google Gemma, and availability varies by region.
To use a model that isn't hosted for you, download a base model from the Databricks Marketplace or Hugging Face and register it in Unity Catalog. Provisioned throughput supports base models of supported architectures, fine-tuned variants of them, and custom weights that use a supported base architecture. Databricks recommends the instruct versions of the models in system.ai for deployment and the base versions for fine-tuning.

How does Unity Gateway fit?

Unity Gateway connects agents and applications to models through one API. It routes requests, limits AI spend, manages fallbacks and monitors usage. Access is controlled with Unity Catalog permissions, and external model providers connect with centrally managed credentials. To query a model through Unity Gateway, check for the corresponding model service in your workspace. The Databricks release notes list general availability on August 4, 2026.

FAQs

Can I use open models commercially?

Check the license for each model. gpt-oss uses Apache 2.0. The Llama 4 Community License grants a non-exclusive, royalty-free limited license, requires "Built with Llama" attribution when you distribute, requires a separate license from Meta above a user threshold, and incorporates an Acceptable Use Policy. Databricks notes that customers are responsible for compliance with applicable model terms.

What is a gated model on Hugging Face?

A model whose authors enabled access requests. You agree to share your username and email address with them to access the files, and approval can be automatic or manual.

Can I run a Hugging Face model on Databricks?

If its architecture is supported, register it in Unity Catalog and serve it with provisioned throughput. Check the supported model architectures first.

Are all Databricks-hosted open models available in every region?

No. Availability depends on region, cross-geography settings and available serving capacity, so check the model availability by region page.