GitHub Copilot local sandboxing is generally available. It runs Copilot commands in an isolated environment. Access to files, networks, system capabilities, and credentials is controlled for granular control. It also provides benefits for enterprise teams.
Key points
Availability: Generally available release of local sandboxing for GitHub Copilot.
Isolation: Runs commands with controlled access to files, networks, system capabilities, and credentials.
Sources
GitHub Copilot adds local models and sandboxed tools to Windows
GitHub Copilot is adding support for local models and sandboxed tools on Windows. It will soon decide whether a task should run on-device or use cloud‑scale models.
Key points
Feature: Local models and sandboxed tools will be available on Windows.
Function: Copilot will choose between on-device intelligence and cloud models for each task.
Sources
- Original post
- Linked resource - Linked in the post
GitHub Copilot to add intelligent local model routing
GitHub Copilot will soon include intelligent local model routing. The feature will automatically route tasks to a local model when it is most suitable. It is presented as part of the Project HydraFusion vision and aims to help users save on AI credits.
Key points
Announcement: Intelligent local model routing will be added to GitHub Copilot.
Goal: The routing will help users save on AI credits.
Sources
GitHub Copilot adds local model support with MAI Code 1.1 Flash
GitHub Copilot will automatically choose between on‑device and cloud models and also allow explicit selection of local models. The on‑device model is MAI Code 1.1 Flash, a 137 billion‑parameter mixture‑of‑experts model quantized to 53 GB. Performance numbers are given for different context lengths.
Key points
Auto orchestration will decide when to run tasks locally or in the cloud, launching at the end of the month.
MAI Code 1.1 Flash uses 75.5 GB peak memory at 256 k context and processes 923.5 tps at 64 k context and 769.8 tps at 128 k context.
Sources
Antigravity integrates StitchByGoogle MCP for native Android builds
Antigravity enables building and testing native Android apps. It connects the StitchByGoogle MCP and the Android CLI. The agent imports UI designs, converts them to Jetpack Compose components, validates in an emulator, and runs the final build.
Key points
Integration: Antigravity links StitchByGoogle MCP with Android CLI.
Workflow: UI designs are turned into Jetpack Compose components and validated in an emulator before final build.
Sources
First TPUs to be launched into orbit
The announcement states that certain hardware tests require space conditions such as solar events, cosmic rays, and radiation‑induced bitflip errors. Because of this, the team is sending its first TPUs into orbit.
Key points
Test environment: hardware can only be tested in space for solar events, cosmic rays, or radiation‑caused bitflip errors.
Deployment: the first TPUs are being put in orbit.
Sources
Project Suncatcher prototype satellite test in orbit
Project Suncatcher is a Google moonshot to test AI hardware in space. A prototype satellite will launch on a SpaceX rideshare mission to evaluate TPU performance under launch vibration, radiation, and thermal conditions. The team has completed vibration testing, radiation testing at UC Davis, and thermal vacuum testing of cooling systems. Results will inform future satellite designs and inter‑satellite laser links.
Key points
Vibration test: satellite shaken on three axes; TPU chips survived forces up to 50‑100 g.
Radiation test: Trillium TPUs endured a total ionizing dose greater than a five‑year mission in a proton beam at UC Davis.