Ecosystem

What is changing?

We are excited to share that we have shipped a new UI for users who read through our changelogs on the GitHub blog webpage! This is meant to make your reading journey a lot more seamless by providing necessary contextual elements around each post.

What exactly are the changes?

• The overall look and feel of the UI now closely mirrors the experience that you get when you log into github.com.
• You’ll see that new posts are categorized into one of three categories: New Release, Improvement, or Deprecation.
• There is a new set of 12 curated categories that you can filter on to see changes by that product area.
• We have introduced a new summarization feature to help assist quick consumption of content so you don’t have to read through each individual post (unless you want to!).
• Posts are now grouped into months for easy viewing and archived into years for last referencing.

Why is it important?

We heard your feedback that it can be hard to read through the latest changes and understand what changes are new features releases versus what are slight improvements to existing feature sets. We hope that this will make it clearer, and you will be able to filter on the curated, new feature categories that are most relevant for you. This will also enable you to catch up on all the changes of that category quicker, supported by our new summarization feature.

When is it changing?

The aforementioned changes will be in effect starting April 23, 2025. You may have already experienced this in its newest form!

MAI-DS-R1 is now available on GitHub Models.

MAI-DS-R1 is an updated version of DeepSeek-R1, refined by Microsoft AI. It handles complex queries more effectively, works across multiple languages, and provides access to previously restricted information. The model maintains the reasoning strengths of the original while improving reliability.

Try, compare, and implement this model in your code for free in the playground or through the GitHub API. Compare it to other models using side-by-side comparisons in GitHub Models.

To learn more about GitHub Models, check out the docs. You can also join our community discussions.

The latest AI models from Cohere, Command A and Embed 4, are now available on GitHub Models.

Command A is a multilingual model designed for business-critical applications like retrieval-augmented generation (RAG) and agentic tasks. It excels at supporting knowledge assistants, improving demand forecasting, and optimizing eCommerce search.

Embed 4 is a multilingual model that transforms text, images, and mixed formats into unified vector representations. It is well-suited for processing high-resolution images and extracting key details from files like PDFs, slides, and tables.

Try, compare, and implement Command A in your code for free in the playground. Both Command A and Embed 4 are also available through the GitHub API for seamless integration into your applications.

To learn more about GitHub Models, check out the docs. You can also join our community discussions.

Alongside the launch of GPT-4.1 in GitHub Models, we’re introducing GPT-4.1-mini and GPT-4.1-nano—lightweight variants of OpenAI’s latest model. Designed for high performance with lower cost and latency, these models are ideal for real-time applications and workloads that involve parallel or chained model calls.

Both inherit the core strengths of the GPT-4.1 series, including enhanced coding capabilities, improved instruction following, long-context understanding, and multimodal support (text and image). With features like parallel function calling and structured output generation, GitHub Models users can now choose the right-sized model for their specific needs—whether building chatbots, coding copilots, or AI-powered agents.

• GPT-4.1-mini: Combines strong general-purpose reasoning with low cost and latency, supporting both text and vision use cases.

• GPT-4.1-nano: Offers even lower cost and latency, ideal for lightweight tasks and high-frequency usage at scale.

Try, compare, and implement these models in your code for free in the playground (GPT-4.1-mini and GPT-4.1-nano) or through the GitHub API.

To learn more, visit the GitHub Models documentation, and join the community discussions to share feedback and connect with other developers.

You can now use the built-in GITHUB_TOKEN from GitHub Actions to authenticate requests to GitHub Models. This simplifies your workflows by integrating AI capabilities directly into your actions, eliminating the need to generate and manage Personal Access Tokens (PATs).

With this update, creating and sharing AI-driven GitHub Actions has never been easier. Add AI to your workflows effortlessly, whether it’s generating issue comments or reviewing pull requests.

Try it out today and streamline your automation with integrated AI.

GitHub Models empowers every developer to effortlessly incorporate AI into their GitHub workflows.

For more details, check out our documentation or join our community discussions.

The latest AI models from Meta, Llama-4-Scout-17B-16E-Instruct and Llama-4-Maverick-17B-128E-Instruct-FP8, are now available on GitHub Models.

Llama-4-Scout-17B is a 17B parameter Mixture-of-Experts (MOE) model optimized for tasks like summarization, personalization, and reasoning. Its ability to handle extensive context makes it well-suited for tasks that require complex and detailed reasoning.

Llama-4-Maverick-17B is a 17B parameter Mixture-of-Experts (MOE) model designed for high-quality chat, creative writing, and precise image analysis. With its conversational fine-tuning and support for text and image understanding, Maverick is ideal for creating AI assitants and applications.

Try, compare, and implement these models in your code for free in the playground (Llama-4-Scout-17B-16E-Instruct and Llama-4-Maverick-17B-128E-Instruct-FP8) or through the GitHub API.

To learn more about GitHub Models, check out the docs. You can also join our community discussions.

DeepSeek-V3-0324 is now available on GitHub Models.

DeepSeek-V3-0324 is a 671B parameter Mixture-of-Experts (MoE) model that builds notable updates on top of its predecessor, DeepSeek-V3. These include enhanced reasoning capabilities and improved function calling accuracy. This model also excels in Chinese writing proficiency and includes advanced search capabilities for Chinese.

Note: DeepSeek-V3 will be deprecated on Friday, April 11th, 2025. We recommend transitioning to DeepSeek-V3-0324 to take full advantage of its enhanced features.

Try, compare, and implement this model in your code for free in the playground or through the GitHub API. Compare it to other models using side-by-side comparisons in GitHub Models.

To learn more about GitHub Models, check out the docs. You can also join our community discussions.

Mistral Small 3.1 (25.03) is now available in GitHub Models.

Mistral Small 3.1 (25.03) is a versatile AI model designed to assist with programming, mathematical reasoning, dialogue, and in-depth document comprehension. Equipped with multimodal capabilities, it processes both text and visual inputs, making it suitable for chat-based interactions and instruction-following tasks.

Try, compare, and implement this model in your code for free in the playground or through the GitHub API. Compare it to other models using side-by-side comparisons in GitHub Models.

To learn more about GitHub Models, check out the docs. You can also join our Community discussions.

Fine-grained Personal Access Tokens (PATs) have been used by millions of users to make tens of billions of API calls over the last two years in public preview. In that time, we’ve added requested features such as management APIs and webhooks, mandatory expiration policies, and usability improvements.

However, feedback has been clear on one item in particular – while fine-grained PATs solve a significant set of challenges in their current state, many organizations cannot fully adopt them due to the lack of support statements and the risk of breaking changes while they’re in public preview. Our goal at GitHub is to ensure that everyone can secure their workflows as best they can, which is why we’re graduating fine-grained PATs to a generally available (GA) state.

Changes with this release

This update brings two major changes to PATs at GitHub. Most notably, fine-grained PATs are now enabled by default for all organizations on GitHub, unless that organization or enterprise explicitly disabled them during the preview. The PAT approval flow is also enabled by default, so developers must request organization owner approval in order to successfully use their fine-grained PAT against their organizations.

We’re also updating the release state for both fine-grained PATs and PAT expiration policies. These features are now fully supported by GitHub and adhere to the same breaking change policies as the rest of the product. While there are some scenarios where fine-grained PATs are not yet supported, your organization should be confident in suggesting, or even requiring, the use of these more secure tokens.

Administrators, auditors, and security teams can also look for improved auditability of PATs – the token_id is now included in all API calls and supported as a built-in filter in the audit logs. With this filter, you can now easily track the use of a token throughout your enterprise or organization.

Customers on GHES should expect these changes to arrive in version 3.17.

Feature gaps in fine-grained PATs

There are several scenarios where fine-grained PATs are not a suitable solution at this time. GitHub continues to invest in building more secure access patterns and will implement these capabilities over time. You can track our progress and goals on our public roadmap. The most notable scenarios are:

• Calling APIs that manage the Enterprise object (e.g. SCIM APIs or creating organizations)
• Accessing multiple organizations with a single token
• Contributing to repositories where you’re an outside collaborator or an unaffiliated open source contributor
• Accessing internal repositories in your enterprise, outside of a targeted organization
• Calling the Packages and Checks APIs

We’re currently focused on implementing enterprise access for GitHub Apps and fine-grained PATs so that enterprise owners can reduce the over-permissioning of their current automation solutions. After that, we’ll continue to invest in this area with a goal of enabling organizations to eventually disable the use of PATs (Classic) for their resources.

To learn more about fine-grained PATs and how your organization can control them, see our documentation on managing your personal access tokens, and enforcing policies for PATs in your enterprise.