Where are the models hosted?

[StefMa]

Select Topic Area

Question

Body

GitHub Copilot has a good overview where the used models are hosted.
Like Claude on AWS, PBC, or GCP. Or Gemini on GCP, and GPT on Azure.

It would be really helpful to have a similar list in the documentation showing where each model (from GitHub Models) is hosted. This kind of transparency would make it easier for companies to evaluate which models they can use in line with internal policies around data privacy, security, and compliance.

For example, at my company, we’re allowed to use models hosted on Azure or GitHub, but not those on AWS or GCP.

[LukasBel]

GitHub Copilot Model Hosting Locations

This document outlines where the different AI models accessible through GitHub Copilot are hosted. Understanding this is important for compliance with organizational policies regarding data processing and infrastructure usage.

Core GitHub Copilot Infrastructure & Default Models (OpenAI)

• The primary infrastructure for GitHub Copilot, including its proxy services and the large language models (LLMs) it relies on (developed by GitHub, OpenAI, and Microsoft), is hosted within Microsoft Azure tenants owned by GitHub.
• This means the default code completion features and likely the default OpenAI models offered through Copilot Chat (like versions of GPT-4, GPT-4o, etc.) run on Azure.
• Data transmitted to these core services is encrypted in transit (TLS) and at rest. GitHub states that prompts and suggestions for code completion in the editor are typically discarded after a response is generated and are not used to train the foundational public models unless users opt-in (for specific plans/features).

Alternative/Partner Models in Copilot Chat

GitHub Copilot Chat allows users (often requiring explicit opt-in or administrator approval) to select alternative AI models. The hosting for these models varies:

• Claude Models (e.g., Sonnet 3.5, Sonnet 3.7):

  • When used via GitHub Copilot, these models are explicitly hosted by Amazon Web Services (AWS), Anthropic PBC, and Google Cloud Platform (GCP).
  • GitHub has agreements with these providers to ensure data is not used for their model training (e.g., zero data retention with Anthropic, specific commitments from AWS Bedrock and Google Cloud).

• Gemini Models (e.g., 1.5 Pro, 2.0 Flash, 2.5 Pro):

  • These models are provided through a partnership with Google Cloud and are hosted on GCP.

• Other OpenAI Models (e.g., o1, o3, o3-mini, o4-mini, GPT-4.1):

  • While accessed via Copilot, these are OpenAI models. Given the core infrastructure, they are most likely served from the Azure environment managed by GitHub/Microsoft.

Note: Organizations using GitHub Copilot Business or Enterprise can often configure policies to restrict or allow the use of specific models based on their hosting provider to maintain compliance.

[StefMa]

Hi @LukasBel,

Thank you very much for the detailed overview.

I believe I may not have expressed my question clearly enough earlier.
I'm not referring to the hosting of GitHub Copilot LLMs, but rather to the GitHub Models LLMs, which can be found here: https://github.com/marketplace/models.

Would you happen to have any information regarding the hosting details for those?

Thanks again!

[LukasBel]

The Only Github Marketplace Models I Could Find Data On In Regards To Their Hosting:

• OpenAI Models (e.g., GPT-4o, GPT-4.1, o3, o4-mini)

  • Hosting: Primarily hosted on Microsoft Azure.
  • Notes: These models are often developed in close partnership with Microsoft and GitHub, utilizing Azure infrastructure.

• Anthropic Models (e.g., Claude 3.5 Sonnet, Claude 3.7 Sonnet)

  • Hosting: Hosted on Amazon Web Services (AWS), Anthropic PBC's infrastructure, and Google Cloud Platform (GCP).
  • Notes: GitHub accesses these models via agreements with the providers, ensuring specific data handling protocols (zero data retention agreements with Anthropic PBC).

• Google Models (e.g., Gemini 1.5 Pro, Gemini 2.0 Flash, Gemini 2.5 Pro)

  • Hosting: Hosted on Google Cloud Platform (GCP).
  • Notes: Provided via a direct partnership between GitHub and Google Cloud.

• Microsoft Models (e.g., Phi series)

  • Hosting: Primarily hosted on Microsoft Azure.

• Meta Models (Llama series)

  • Hosting: The specific hosting infrastructure when accessed via GitHub Models is less explicitly documented in public materials. It likely involves Meta's own infrastructure or a major cloud partner (Azure, AWS, GCP) accessed through GitHub's agreements.

• Cohere Models

  • Hosting: Similar to Meta models, hosting when accessed via GitHub Models likely utilizes Cohere's platform or a partner cloud (Azure, AWS, GCP) managed through GitHub's integration agreements.

• DeepSeek Models

  • Hosting: As these are often open models, hosting via GitHub Models likely uses infrastructure on a major cloud partner (Azure, AWS, GCP) accessed via GitHub's platform agreements.

[KateCatlin]

Hi @StefMa! Thanks for the question. At this time, all of our models are hosted on Azure via GitHub, and those are also the only places your data flows through. This may change in the future, so I'll make a note to update docs with more clarity there.

[chb-ioki]

Hi @KateCatlin,
thank you for the explanation!

Just one follow-up question for clarification: When you say all of our models are hosted on Azure via GitHub, does this include i.e. Claude?

[KateCatlin]

Hey @chb-ioki, thanks for asking! I think there's an important distinction to be made here! When I talk about Models, I'm referring to GitHub Models, a catalog of models you can call as a service in your AI-powered apps: https://github.com/marketplace/models

Our catalog does not include Claude.

The models you use within Copilot to write code (which does include Claude) are a separate entity and business model and you'd have to go ask in the Copilot community forum about where those are hosted!

[chb-ioki]

Aaah, alright. Thank you for the hint!