Intelligent local model routing is coming to GitHub Copilot
Read the Portuguese translation →
Author: GitHub · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Intelligent local model routing is coming to GitHub Copilot · Channel: GitHub · 02:09 · Watch on YouTube
Topics: Task and workflow automation, AI developer tools, Model inference and optimization, Local and open-source models
Summary
Section titled “Summary”GitHub announces intelligent local model routing for Copilot, extending its model orchestration beyond cloud models. Supported local models will be discovered automatically, and capable devices will be able to install MAI Code 1.1 Flash from Copilot. With auto-routing, Copilot can use local models for suitable tasks, automations, and offline work, without AI-credit costs. Project HydraFusion is available today in experimental settings, while local routing is coming soon.
Key points
Section titled “Key points”- Project HydraFusion orchestrates work across cloud models and can reduce costs by up to 67%.
- GitHub’s vision is to make unlimited, unmetered intelligence available on every desktop.
- Copilot will automatically discover models installed through supported providers such as Microsoft Foundry Local and Ollama.
- Capable devices will be able to install MAI Code 1.1 Flash directly from GitHub Copilot.
- When auto-routing is enabled, Copilot can send simple tasks to local models.
- Users can plan with frontier models such as Astra and hand off implementation to a local model.
- Local models can run scheduled background automations without using AI credits.
- Local models remain available offline.
- Project HydraFusion is available now in experimental settings; intelligent local model routing is coming soon.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | HydraFusion and the vision for local intelligence | GitHub introduces HydraFusion’s cloud model orchestration and explains its ambition to bring unmetered intelligence to desktops. |
| 00:33 | Automatic discovery and local model installation | Copilot will discover models from supported providers and offer MAI Code 1.1 Flash for installation on capable devices. |
| 01:06 | Auto-routing for everyday tasks | Copilot can route simple tasks to local models or hand off implementation after planning with a frontier model. |
| 01:37 | Automations, offline use, and availability | Local models can power background automations without AI-credit costs and work offline, while HydraFusion is available in experimental settings and local routing is coming soon. |
Quotes
Section titled “Quotes”unlimited and unmetered intelligence belongs on every desktop. — 00:00
Any models that you have installed with supported providers, such as Microsoft Foundry Local or Ollama, will just automatically be discovered — 00:33
the inference is completely free for you. — 01:06
costing you exactly zero AI credits to use. — 01:37
zero dollars and zero cents. — 02:09
Concepts and entities
Section titled “Concepts and entities”GitHub Copilot, Project HydraFusion, Microsoft Foundry Local, Ollama, MAI Code 1.1 Flash, Astra, roteamento automático, modelos locais
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Intelligent local model routing is coming to GitHub Copilot. A few weeks ago, we announced Project HydraFusion, an intelligent model orchestration that plans, implements, and executes across a wide range of cloud models, giving you frontier-level performance while significantly reducing cost, up to 67%. But our vision goes beyond just the models that are available in the cloud, because we think that unlimited and unmetered intelligence belongs on every desktop.
00:33 This is why we’re pleased to announce that soon, Copilot will intelligently route to and leverage local models on capable devices. First, you’ll no longer have to configure local models in GitHub Copilot itself. Any models that you have installed with supported providers, such as Microsoft Foundry Local or Ollama, will just automatically be discovered and will appear in the model picker. Capable devices will also be able to install MAI Code 1.1 Flash, a local model available for install right from GitHub Copilot.
01:06 Copilot will then automatically use these models when set to auto-routing. There are a myriad of scenarios where local models are perfect, and the inference is completely free for you. For instance, when set to auto-routing, Copilot can automatically offload to a local model for simple tasks, such as explaining how a codebase works or how to run a project. You can also plan with large frontier models, such as Astra, and then hand off to a local model to implement while you sleep.
01:37 And users love automations in GitHub Copilot. These are tasks that run on a schedule in the background. Local models work incredibly here, able to work on very long-running tasks and costing you exactly zero AI credits to use. You can also go completely offline and still have access to your local models, which is perfect for working disconnected or when you just don’t have an internet connection. Project HydraFusion is available today in GitHub Copilot.
02:09 You can look for it under experimental settings. And capable devices will be able to use intelligent local model routing very soon, including MAI Code 1.1 Flash. With GitHub Copilot, you get intelligent multi-model routing, both in the cloud and very soon, with local models that run on your machine and cost you exactly zero dollars and zero cents.