Google has released Android Studio Quail 4 to the stable channel, turning several AI-assisted development features from the Quail preview cycle into production tooling. Android Developers announced the release on September 1, 2026. The most consequential changes are bundled Android-specific skills, native management of local Gemma 4 models and a clearer interface for supervising parallel agent sessions.
Android knowledge is packaged as skills
Quail 4 ships with 23 curated skills created by the Android team. They cover tasks such as Android Gradle Plugin 9 upgrades, profiling, Navigation 3 migration and adaptive-interface work. Android Studio examines a prompt and uses installed skill metadata to invoke relevant guidance, so the coding agent receives task-specific instructions instead of relying only on generic model knowledge.
That distinction matters because Android APIs, build tooling and migration procedures change frequently. A skill can give the model a maintained procedure and constraints, but it does not make generated code automatically correct. Developers still need to inspect diffs, run tests and verify behavior on the devices and API levels they support. Teams can disable the bundled skills in settings, create their own, or install the same Android skills for compatible command-line AI tools through Android CLI.
Gemma 4 can run inside the IDE
Android Studio now integrates Gemma 4 as a locally managed model provider. The IDE downloads, verifies and updates the selected weights and includes an inference engine, removing the need to configure a separate local-model server for this workflow. Google says the smallest models can run with 12 GB of RAM, while systems with at least 32 GB should provide a better experience.
Local execution is useful when source code must stay on the workstation, internet access is unreliable or cloud quotas are a constraint. The core local workflow does not require an API key and can operate offline. Those properties describe where inference happens; they are not a substitute for a project's own security review, access controls or dependency policy. Hardware capacity will also shape model size and responsiveness.
Parallel work gets more visible controls
Quail 4 adds status indicators for background chats, links recognized code symbols and paths in responses, and consolidates task and walkthrough artifacts into a Summary of Changes view. The interface shows when an agent is working, waiting for input or ready for review. These are supervision improvements rather than evidence that parallel agents will finish a task correctly.
The practical benefit is easier separation of work: a developer can keep a refactor, investigation or documentation task in its own conversation and see which session needs attention. The final diff remains the important review surface.
What developers should check before adopting it
Teams moving to Quail 4 should first confirm plugin and build compatibility, then test bundled skills on representative tasks in a branch. For local Gemma 4, measure memory pressure and response time on the actual developer hardware. Organizations with code-handling rules should document which model provider is selected and whether any optional cloud features are enabled.
The release makes AI assistance more configurable: Android-specific instructions, a local-model path and cloud providers can coexist in one IDE. The useful question is therefore not whether to enable every feature, but which execution path fits each repository's privacy, review and reliability requirements.