OneLLMONELLM · Guides
Local AI on iPhone: downloads, computer servers and cloud models.
Start by choosing where inference runs: on your iPhone, on a computer you connect to, or through a cloud API. Ollama, LM Studio, Hugging Face and ModelScope serve different roles from model families such as Qwen, Gemma and MiniMax. A model name alone is not enough to choose a setup.
Ollama and LM Studio: computer inference, phone access
Both tools can serve models from a computer. In that setup, the computer handles the main inference workload and memory requirements. The phone needs a reachable service address, and the computer must remain running.
Swipe horizontally to view the full table.
| Option | Where the model runs | Relationship to OneLLM |
|---|---|---|
| On-device model | Downloaded to an iPhone or iPad for device inference. | Choose a compatible file from the app’s available model catalog. Inference can work offline after download. |
| Ollama service | Handled by the computer service; check whether the selected model is local or cloud-backed. | OneLLM provides an Ollama connection. Use a reachable computer address and confirm the model list loads. |
| LM Studio service | A model loaded on the computer, exposed through local or network APIs. | Its official APIs include OpenAI-compatible endpoints. Direct OneLLM compatibility is not verified here; an LM Studio address is not interchangeable with an Ollama address. |
On your phone, localhost refers to the phone itself, not your computer. Use the correct LAN address or a configured private-network address, and check the computer’s firewall and service binding. There is no need to expose the model service directly to the public internet. See the Ollama API documentation and LM Studio server guide for service configuration.
Hugging Face and ModelScope: where model files come from
A download source is separate from the execution environment. OneLLM’s Hugging Face browser finds GGUF files. Some catalog models offer Hugging Face and ModelScope download sources. Access to a repository does not establish that every file in it can run on your phone.
- Read the complete model name.Check the version, parameter size, quantization and file purpose rather than only the family name.
- Check device resources.Downloads use storage, while inference also requires memory. File size is not the same as runtime memory use.
- Check format and components.GGUF is a file format; compatibility also depends on architecture and the app’s runtime. Multimodal models may require extra vision components. A text-model download alone does not establish image understanding.
- Follow access requirements.For gated repositories, complete the platform’s authorization or terms requirements. If a download fails, check access, connectivity and free storage.
Read the Hugging Face GGUF guide for format details. ModelScope availability follows the download options for each model in the app; this is not a claim that the entire ModelScope catalog is integrated.
Qwen, Gemma and MiniMax: local files versus cloud access
Swipe horizontally to view the full table.
| Name | How to interpret it in OneLLM | What to check |
|---|---|---|
| Qwen | The catalog includes selected local Qwen entries, alongside a separate Qwen cloud-provider option. | For local use, check the file and device memory. For cloud use, check the account region, model access and usage charges. |
| Gemma | The catalog includes selected local entries, such as Gemma 3 270M. A family name does not imply compatibility with every release. | Start with models listed in your installed version and verify architecture, format, size and usage terms. |
| MiniMax | This guide covers the MiniMax cloud API option in OneLLM, not an on-phone offline feature. | Use your own account and API key, select the appropriate region and model, and account for network access and provider charges. |
Catalog entries and options can change between app versions. Refer to the Qwen official site, Gemma documentation and MiniMax API prerequisites. OneLLM is an independent client, not a representative of these brands.
Check your first setup in this order
- Choose on-device, computer-service or cloud access. Decide whether you need offline inference or remote access.
- Start with a smaller compatible model or a working service and send a short prompt.
- Review the response, initial wait and device behavior before increasing context or task complexity.
- For work on computer code projects, see the OpenCode mobile guide. Model chat and coding-agent conversations use different entry points.