Ollama is a free, open-source model runner from Ollama Inc. for Windows, macOS, and Linux. It downloads and runs AI models locally, provides chat and developer interfaces, and connects applications to model responses. Optional cloud services run models on remote servers and add paid usage options.
Local models or cloud execution
Users choose a model from the library and run it on their computer, or select a cloud model. Local execution downloads model weights before generating responses. Cloud execution sends requests to Ollama’s servers without downloading those weights.
The app and command line use account sign-in for cloud access. Applications can also call the hosted API directly with an API key. Local API requests need no key.
| Cloud account | Scope |
|---|---|
| Free | Starter models and usage; additional credits unlock all cloud models. |
| Pro and Max | Included usage credits and concurrent requests; Max adds early model access. |
| Team | Shared usage credits with centralized billing and administration. |
| Enterprise | Model access controls and budgets for users and API keys. |
Cloud plans impose concurrency limits. Excess requests enter a queue, and a full queue rejects further requests. These account limits do not meter local model execution.
Chatting and managing models
The Windows and macOS apps provide model downloads and chat. Users can drag text, PDF, and code files into a conversation for questions about their contents. A vision-capable model also accepts images alongside text. Context settings control how much material the model can consider.
In the terminal, the run command opens a conversation with a model. It accepts ordinary prompts, multiline input, or input from another program. Follow-up messages continue the interaction until the user exits.
Separate commands download models, list the local collection, and show which models currently run. Users can stop a running model without deleting it, or remove a model from the collection.
Defining a custom model
A Modelfile combines a base model with instructions and generation settings. Users choose a system message, prompt template, conversation examples, and parameters such as temperature or stop sequences. The create command turns that definition into a model with its own name.
Ollama can also import compatible Safetensors weights or GGUF files through a Modelfile. Users can publish a model under their account name so other people can download and run it.
Application connections
The API accepts a model name and conversation messages, then returns the next response. Applications can receive reply chunks as the model generates them or request one complete response. Official Python and JavaScript libraries expose these operations.
Compatibility endpoints support subsets of the OpenAI and Anthropic APIs. The launch command configures external coding tools to use an Ollama model and can start the selected tool. On macOS, the Apps view also connects supported desktop assistants.
Tool requests and web content
Tool calling lets an application supply functions that a model may request. The application executes each request, returns the result, and asks the model to continue. A loop can repeat this exchange until the model produces its answer.
Ollama’s web APIs add search results or page content to this process. Search accepts a query and returns titles, URLs, and relevant text. Fetch accepts a page URL and returns its content and links. These services require an account and API key.
Data beyond a chat reply
- Structured outputs constrain a response to JSON or a particular JSON schema. Developers provide the fields and types they expect. Ollama’s cloud does not support this capability.
- Embedding models convert individual texts or batches into numeric vectors. Applications can use those vectors for similarity searches and document retrieval.
- Decision models accept text or structured state with questions and criteria. They return labels, yes-or-no probabilities, or scores across an ordered scale.
- Thinking-capable models separate their reasoning output from their final answer. Controls vary by model, with on/off choices or named levels.
Experimental image generation
Experimental image generation currently supports macOS only. A text prompt produces an image file in the current directory. Users can adjust width, height, generation steps, random seed, and negative prompts. Compatible terminals preview the resulting image inline.
Speech from text
On Apple silicon Macs, speech models generate audio from text or a chat model’s reply. The command line plays audio as it arrives, or writes a WAV file when users redirect output. The speech API offers WAV or raw PCM audio. Speech models handle audio generation rather than chat or embeddings.
