Every step a phone agent takes is a model call. Read the screen, pick one action, tap, repeat. A simple task can be 10 of those. A test suite can be thousands.
So the first question people ask after getting Mobilerun working is always the same: can I run this on something cheaper? Or on my own machine, with no API key at all?
Yes. The open-source Mobilerun framework doesn't care where the model lives. This post shows what runs where, the exact local setup with Ollama and Qwen, the hosted setup for DeepSeek and OpenRouter, and what actually changes your results.
The problem with "which model is best?"
You'll find plenty of opinions about which model is best for phone agents. You won't find a published Mobilerun benchmark broken out by Ollama, DeepSeek, Qwen and OpenRouter. We haven't published one yet either.
That matters, because the answer depends on things only you know: your tasks, your device, your hardware, and whether you run with vision. So instead of a ranking, we'll give you the setup for each option and a simple way to measure them on your own work.
What runs where
The Mobilerun framework runs on your machine. It controls an Android device through Portal and ADB. The LLM provider is a separate choice you make in config or on the command line.
Here's how the options split:
- Ollama runs inference locally. No model API key.
- DeepSeek is a hosted API. It uses DEEPSEEK_API_KEY.
- OpenRouter is a hosted API that routes to many models. It uses OPENROUTER_API_KEY.
- Qwen is a model family, not a provider. You can run Qwen locally through Ollama, and Qwen models are also available through hosted providers like OpenRouter.
Install and run it locally with Ollama and Qwen
You need Python 3.11 to 3.13. Python 3.14 isn't supported yet, and the docs warn that installing on 3.14 silently falls back to an old version. You also need ADB and an Android device with USB debugging on.
Install the framework:
Install Ollama, connect your Android device, then set up Portal and check the connection.
mobilerun setupmobilerun pingCopy
If ping prints "Portal is installed and accessible. You're good to go!", you're ready. Pull the model and run a task:
ollama pull qwen3:8bmobilerun run "Turn on dark mode" --provider Ollama --model qwen3:8b --base_url http://localhost:11434Copy
That's the whole loop running on your hardware. The framework talks to Ollama at localhost:11434, and no request leaves your machine for inference.
Pass --provider and --model together. The CLI guide is clear that passing only one makes the run fail before the agent starts.
To make Ollama the default instead of typing flags every time, add a profile to your config file. Mobilerun creates it at ~/.config/droidrun/config.yaml on Linux, ~/Library/Application Support/droidrun/config.yaml on macOS, and %LOCALAPPDATA%\droidrun\droidrun\config.yaml on Windows. Yes, the folder still says droidrun. You can also point MOBILERUN_CONFIG or --config at your own file.
One Ollama detail worth knowing: Mobilerun sets the context size itself. The default context_window is 32,768, which keeps most models on the GPU. Setting it to -1 asks for the model's full context, which preallocates the whole KV cache and can spill onto your CPU.
DeepSeek and OpenRouter: hosted, pay per call
For the hosted options, set the key and pick the provider and model with CLI flags.
DeepSeek:
export DEEPSEEK_API_KEY=your-keymobilerun run "Turn on dark mode" --provider DeepSeek --model deepseek-flashCopy
OpenRouter, here with a hosted Qwen model:
export OPENROUTER_API_KEY=your-keymobilerun run "Turn on dark mode" --provider OpenRouter --model qwen/qwen3.8-flashCopy
Or save them as profiles so every run uses them:
llm_profiles: fast_agent: provider: DeepSeek model: deepseek-flashllm_profiles: fast_agent: provider: OpenRouter model: qwen/qwen3.8-flashCopy
A note on mobilerun configure. The wizard's provider list today is gemini, openai, anthropic, xai, ollama, openai_like, minimax and zai. DeepSeek and OpenRouter aren't in it as named options, so use the CLI flags or the YAML profile for those. Model IDs change often on both services, so check DeepSeek's and OpenRouter's model lists before you copy ours.
What affects your results
Swapping the model is one line. Whether the task finishes is a different story. These are the four things that change the outcome.
Accessibility tree by default. Out of the box, the agent reads Android's accessibility tree, a structured list of what's on screen with text and bounds. No screenshots. That works with text-only models, including qwen3:8b.
--vision adds screenshots. With vision on, the agent gets a screenshot alongside the tree on every step. That needs a vision-capable model. If you test vision on a local model, pull one that accepts images. For local vision models that downsize images in a way Mobilerun can't detect, the docs suggest capping the screenshot so tap coordinates stay accurate:
agent: model_screenshot_max_side: 1280Copy
There's also --vision-only, which drops the tree and drives the device from screenshots alone.
Native tool calling vs the fallback protocol. Mobilerun uses native tool calling where the model supports it. For other models, it falls back to a text protocol where the model writes tool calls as XML and the framework parses them. In v0.7.4 the native path covers providers like DeepSeek, OpenAI and xAI. Ollama and OpenRouter go through the fallback. If a model writes broken tool-call markup three times in a row, the run stops, and smaller models hit this more often.
Your hardware. With Ollama, speed is your GPU, your RAM and the model size. The same model can feel quick on one laptop and painful on another.
Success and speed will vary with the model, the task, the device and the hardware. That's why the next section exists.
Some tips to keep the numbers honest:
- Run mobilerun --version and record it. Results change between releases.
- Keep --steps the same across runs. The default is 15, and complex tasks may need more.
- Turn on --tracing to see tokens per step, or use --debug and --save-trajectory action to see where a run went wrong.
- Count completed tasks, not tokens. A cheap model that fails half the time isn't cheap.
On cost: Ollama removes per-call model API fees, but local hardware still has a cost. The GPU, the power and the machine that's busy while it runs all count. Hosted APIs flip that. No hardware to own, but you pay for every call.
Honest limits
- There are no published per-provider Mobilerun benchmarks yet. Anyone giving you a confident ranking is guessing.
- Small local models are more likely to drift from the tool-call format, especially on busy screens.
- qwen3:8b is a text model. Keep --vision off with it, or switch to a vision-capable model.
- The framework is the open-source, bring-your-own-device path. You manage the phone, ADB, Portal and the model.
Don't want to manage a device?
If you'd rather skip the phone and the ADB setup, Mobilerun Cloud gives you hosted devices and a managed agent behind an API. You send a task, it runs on a phone in the cloud. If you want to keep your own model and your own device, the open-source framework will always let you.
Frequently Asked Questions
Can I run Mobilerun with no API key at all? Is Qwen a provider?
Which model is the cheapest per task? Can I use DeepSeek or OpenRouter from mobilerun configure?
The takeaway
The model is your choice, and the right one is the one that finishes your tasks on your hardware. Measure it, don't guess it. If you'd rather not manage a device at all, start at cloud.mobilerun.ai.
Docs: Quickstart · CLI and providers · Ollama profiles · Vision mode
Also Read:

Top comments (0)