Meta's Reality Labs relies on device emulation to develop and test software for AI glasses and other non-phone Android hardware without depending on scarce physical devices. The quality of that testing hinges on how faithfully a virtual device reproduces the software behavior of the real product. Lightweight emulators such as Goldfish cover only part of the way, and the gaps between emulated and real behavior can let bugs survive until late in development — while also eroding trust in emulation itself and pushing teams back toward physical hardware.
Cuttlefish is AOSP's virtual-device framework for testing Android software against virtual hardware. It boots the same system images as real hardware and runs a realistic software stack on top of virtual hardware, so it tracks physical-device behavior more closely than a lightweight emulator. The catch: Cuttlefish is supported only on Linux, while Reality Labs' application and library developers work on Apple silicon Macs. That forced the team to support Goldfish and Cuttlefish side by side.
Why two emulators became untenable
Running both platforms cost more than duplicated engineering time. Tools each needed support for two targets, computing resources and infrastructure were split, and developers had to pick an emulator based on their working setup: Goldfish ran everywhere but with lower fidelity, Cuttlefish offered higher fidelity but required Linux. The AOSP split mirrors the same divide, with Goldfish aimed at application developers and Cuttlefish at Android platform developers.
The stack, and where macOS broke it
Cuttlefish is the orchestration layer; the component that actually runs the guest OS is a virtual machine monitor (VMM). On Linux, Cuttlefish normally uses crosvm, but it previously carried support for QEMU as an alternative VMM — an integration that had gone inactive over time. Reviving QEMU support is what made native Apple silicon support on macOS possible.
Many of Cuttlefish's virtual devices use vhost-user backends: separate processes that hand virtio devices to QEMU over a socket. The design is clean, but much of it was written with a hard Linux assumption, and three of those assumptions fail on macOS. The vhost-user machinery is gated to Linux and will not build or run; the Apple hypervisor imposes resource limits QEMU did not anticipate; and native display, build, and signing paths were missing entirely.
Porting the vhost-user path
The most structurally significant changes landed in QEMU's vhost-user code. The frontends, libvhost-user, and contrib backends such as vhost-user-blk were all gated to Linux. Relaxing those build restrictions lets the vhost-user frontends and libraries build on macOS as well, so Cuttlefish's device backends can hand off to QEMU on a Mac exactly as they do on Linux. Without this step, the higher-fidelity device model that justifies Cuttlefish would remain unavailable outside Linux.
Fitting inside Hypervisor.framework
On Apple silicon, QEMU accelerates through HVF (Hypervisor.framework) rather than KVM. QEMU's HVF backend did not track enough memory regions for larger Cuttlefish configurations, so its tracking-slot count was increased to let those configurations run correctly.
A native display developers can use
Local application development needs an interactive display so developers can view and interact with the virtual device directly. The team moved to QEMU's native Cocoa display to bypass reliance on the WebRTC interface. The changes include:
- BGRA pixel format support for accurate rendering without visual artifacts on Apple silicon.
- Mouse-click-as-touch, translating pointer actions into touch input for touchscreen devices.
- Custom geometry options, adding the
-gflag tographic_console_initfor display configuration.
The work also enabled the unstable Rutabaga structs needed to match the Rust-side graphics ABI used in gfxstream/rutabaga configurations.
Build and signing on Apple silicon
The remaining work was packaging: initial Darwin build support, building QEMU for arm64 on macOS, and toolchain fixes. macOS builds also required the JIT entitlement along with code-signing and notarization support. On the Cuttlefish side, macOS host support was added so the orchestration layer builds and runs on a Mac instead of assuming a Linux host.
What consolidation buys
A single platform reduced maintenance, cut infrastructure requirements, eliminated redundant tooling, improved fidelity, and — critically — put developers on the exact same Cuttlefish virtual devices used in CI.
Sharing one stack also made it simpler to support capabilities across the board: Bluetooth pairing and connectivity, Wi-Fi, cellular, 3D graphics, over-the-air (OTA) update testing, Fastboot, and snapshotting to save and restore device state in CI and development flows.
The same virtual device also reaches external developers building for Meta's ecosystem, so third-party teams test against the identical Cuttlefish-based platform Reality Labs uses to develop and validate the OS. And with AI agents taking a growing role in Android development, execution and runtime validation matter as much as code generation: virtual devices let engineers and automated agents work concurrently in isolated environments, with instant state resets and repeated test runs that remove dependence on shared physical hardware.
Released to the community
The macOS support is useful beyond Reality Labs' own products, so Meta is sharing the work in two parts:
- qemu-wearables: a Meta-maintained QEMU fork carrying the team's commits.
- cuttlefish-wearables: patch files with instructions for patching Android/AOSP.
Both repositories include build and patch-application instructions. Meta says it is exploring ways to contribute the changes upstream to QEMU and AOSP; until then, developers can use the fork and patch set.



