A new on-ramp to model-driven development
Software creation has historically meant writing, customizing, and deploying code. With the rise of generative AI, a second creative path has emerged: assembling applications around machine learning models. Many developers, however, still lack a straightforward way to reach both open and proprietary models. GitHub Models aims to close that gap for GitHub's 100 million-plus developers.
GitHub Models launches in limited public beta today, giving developers access to a range of models—including Llama 3.1, GPT-4o and GPT-4o mini, Phi 3, and Mistral Large 2—through a built-in playground. From there, the workflow extends into Codespaces and VS Code, and finally to production via Azure AI, which provides responsible AI tooling, enterprise-grade security and data privacy, provisioned throughput, and availability in over 25 Azure regions for select models.
| Professor David J. Malan will be putting GitHub Models to the test in Harvard’s CS50 this fall, to enable students to experiment with AI all the more easily. |
Experiment first, build later
The playground lets developers test prompts and tune model parameters for free, directly on GitHub. This is intended to support the same kind of hands-on experimentation that many developers used to learn coding in the first place. Students, hobbyists, and startups can explore popular models from Meta, Mistral, Azure OpenAI Service, Microsoft, and others without leaving their source code management workflow.
GitHub stresses that prompts and outputs from GitHub Models are not shared with model providers and are not used to train or improve models, consistent with GitHub and Microsoft's privacy and security commitments.
Choosing the right model for the job
Different models suit different workloads. Mistral offers low latency, while GPT-4o excels at real-time multimodal applications involving audio, vision, and text. More complex scenarios might combine multiple models—for example, using an embeddings model for retrieval augmented generation (RAG). GitHub Models provides a catalog of options so developers can compare behavior and cost before committing to an approach, and the platform will expand with additional language and vision models as it moves toward general availability.
From playground to production
The experimentation path does not end in the playground. Codespaces provides a zero-friction environment with sample code for various languages and frameworks, letting developers test inference code before integrating it into their own projects. Once validated, developers can wire the models into their applications directly.
For more structured work, GitHub supports running prompt evaluations in GitHub Actions by piping JSON files through the GitHub Models command in the GitHub CLI. Developers can also build GitHub Copilot Extensions on top of GitHub Models. When an application is ready for production, the transition to Azure AI requires swapping a GitHub personal access token for an Azure subscription and credential.
As an AI startup founder and open source maintainer, GitHub Models enables my team to access and experiment with various LLMs in one place. This streamlines our development and lowers the entry barrier for building AI apps.
Models as part of the GitHub platform
GitHub positions this as a natural extension of its role in AI development. Generative AI project creation on the platform has already exceeded 100K projects in the last year. GitHub Copilot now writes nearly half of the code in files where it is enabled, and GitHub Copilot Workspace is pushing toward coding via natural language for developers of all skill levels.
With GitHub Models, the company is betting that putting model access next to source code, issues, pull requests, and workflows will democratize AI development further. The stated ambition is to grow the developer community to one billion people—roughly 10% of the world's population—and let that scale accelerate progress through open collaboration.
