
Running an AI agent on a smartphone without sending every request to a cloud service is becoming more practical. Buddy, an open-source Android applicationAndroid application, is taking this approach further by allowing users to download AI models and give them the ability to interact with applications on their phones.
Introduced in Buddy 2.3.14 on November 19, 2026, the new on-device AI feature supports four models from Google’s Gemma 4 and Alibaba’s Qwen3.5 families. These models run directly on the phone through the llama.cpp inference engine and can continue working without an internet connection after the required files have been downloaded.
What makes this particularly interesting is that Buddy does more than generate text. Its AI agents can read the screen, tap buttons, enter text and navigate Android applications to complete tasks.
What Is Buddy for Android?
Buddy is an open-source AI assistant developed by Hamdi Ghorbel. It provides a floating chat interface through which users can connect AI agents to their Android devices.
Unlike conventional AI chat applications, Buddy is designed to let an agent interact with the phone rather than simply explain how to perform an action.
For example, a user could ask an agent to open a messaging application, find a conversation and prepare a response. The agent can navigate the interface and enter the message, while Buddy requests approval before allowing sensitive actions such as sending or deleting content.
The application supports several agent integrations, including Claude Code, OpenAI Codex, OpenCode, Google’s Antigravity CLI and Grok. Its newer local-model feature adds another option for people who prefer to run AI inference directly on their phones.
Four AI Models Can Now Run Directly on Android
The main addition in version 2.3.14 is a dedicated section called Models on this phone. Users can access it through Settings, followed by Agents and models.
From there, they can download one of four supported models:
- Gemma 4 E2B: Google’s compact 2-billion-parameter-class model.
- Gemma 4 E4B: A larger 4-billion-parameter-class option.
- Qwen3.5 2B: A compact model from Alibaba’s Qwen family.
- Qwen3.5 4B: A larger model for users with more available device resources.
The models are approximately 1.2GB to 5.2GB in size, depending on the selection. Buddy also downloads the Android inference engine, based on the official llama.cpp project, which occupies approximately 33MB after installation.
Once installed, users can select a local model inside an OpenCode chat under the On this phone option. Buddy starts the model when it is needed and automatically stops it after ten minutes without activity. It can also stop inference when the phone becomes too hot.
How Buddy Uses Smartphone Hardware
Running an AI model locally requires processing power, memory and storage. Buddy addresses this by using different hardware execution options depending on the smartphone.
On supported Qualcomm devices, including Snapdragon 8 Gen 2 and newer processors, the application can use the Neural Processing Unit (NPU). It can also run inference through the GPU or CPU, depending on the available hardware and configuration.
CPU execution is particularly important for users with Google Pixel, Samsung Exynos and MediaTek devices. An initial compatibility problem prevented models from starting on some non-Snapdragon phones, but Buddy 2.3.15, released on November 20, fixed the missing CPU component.
The developer has also published preliminary performance measurements from a OnePlus 12. Using Gemma 4 E2B on the phone’s NPU, the reported prompt-processing speed reached approximately 2,000 tokens per second, while text generation reached around 24 tokens per second.
These figures come from the developer’s own testing rather than independent benchmarking. Actual performance will depend on the phone’s processor, memory, temperature and selected model.
AI Agents Can Control Android Applications
Local AI inference is only one part of Buddy’s functionality. The application connects an AI agent to Android through its Accessibility Service and a local Model Context Protocol (MCP) server.
This connection gives agents access to tools that allow them to:
- Read information displayed on the smartphone screen.
- Tap buttons and interact with application interfaces.
- Enter text into supported fields.
- Scroll through pages and navigate applications.
- Open applications and access available phone tools.
Consider a simple example. Instead of manually opening the calculator, entering figures and returning to an AI application, a user can ask Buddy to perform the calculation. The agent can open the calculator, enter the numbers and retrieve the result.
Buddy also supports notification automation, voice interactions, background operation and sub-agents that can investigate different parts of a task simultaneously.
Its background mode is particularly useful because an agent can work through a separate Android screen while the user continues using the regular interface.
Can Buddy’s AI Agents Work Without the Internet?
Yes. One of the main advantages of the new local-model feature is that inference can continue without an internet connection.
After downloading the model and completing the initial setup, users can switch to airplane mode and continue making supported requests. The selected model processes those requests directly on the smartphone instead of sending them to an external inference service.
However, there is an important distinction between Buddy’s local and cloud-based integrations.
Using Gemma 4 or Qwen3.5 through the on-device option keeps model inference local. Using Claude Code, Codex, Grok or another external AI provider still involves sending the information needed for the request to that provider.
Buddy’s privacy policy explains that screen information can include messages, names, addresses and other personal details when an external agent needs them to complete a task.
Users should therefore check which agent and model they have selected rather than assuming that installing Buddy automatically makes every interaction private.
Small Models Still Have Limitations
Although running an AI model locally is useful, smaller models do not perform as reliably as larger cloud-based systems on complicated tasks.
Buddy’s developer acknowledges that the supported models are better suited to straightforward requests, such as checking battery information, setting a timer or examining something displayed on the screen. Longer workflows involving several applications can produce incorrect actions or responses.
To reduce this problem, Buddy introduced Lite mode. It starts with a shorter instruction prompt and approximately a dozen phone tools, loading additional tools only when necessary. The developer reduced the initial prompt from approximately 24,000 tokens to around 4,000.
Buddy also includes approval controls for sensitive actions. Users can review and approve operations before messages are sent, content is published, purchases are made or information is deleted.
These safeguards are important because an AI agent with access to application interfaces can potentially encounter misleading instructions inside messages or webpages.
Availability and Supported Android Devices
Buddy is free and released under the Apache 2.0 open-source licence. It requires a 64-bit Android smartphone running Android 11 or newer.
The developer distributes the application through GitHub and F-Droid. However, Buddy is not available as a general production release on Google Play. According to the project, its use of Accessibility Services to perform tasks inside other applications was the reason Google rejected its production submission.
The latest verified GitHub release is Buddy 2.3.17, published on November 20, 2026. It adds individual controls for enabling or disabling phone tools and optional read-only access to SMS messages and call logs through background mode.
Interested users can obtain the application from the official Buddy GitHub releases page or check its F-Droid listing. F-Droid releases may arrive later than the GitHub version.
What This Means for Android Users
Buddy demonstrates how local language models can become more useful when connected to tools that perform actual tasks on a smartphone.
Instead of requiring an internet connection for every request, users can download supported models and let them handle certain interactions locally. Combining this with Android application control creates an interesting option for people who want more control over their devices and the information they share with external AI services.
There are still limitations. Smartphone hardware varies, model downloads consume considerable storage, and smaller models can struggle with complicated instructions. Buddy is also an early-stage project, so users should expect occasional errors and review its permissions carefully.
Nevertheless, bringing Gemma 4 and Qwen3.5 into an Android agent environment is a practical demonstration of what locally running AI can do beyond answering questions. The next challenge is making these agents reliable enough to handle longer tasks without requiring constant supervision.
Discover more from Aree Blog
Subscribe now to keep reading and get access to the full archive.


