Privacy & Local AI
Local Models
River works with cloud AI models (Anthropic, OpenAI, xAI, Gemini, DeepSeek) and with local models that run on your own computer through Ollama. You chat with River the same way; the model that answers runs on your machine, not at an AI company.
Your chats are still saved in River's cloud, and each request passes through River's servers on its way to your computer. See How Privacy Works.
When to use local models#
Use local models when:
- Privacy matters. The model that answers runs on your own hardware, and so do the helpers River starts in that chat, which keeps most of a chat there, not all of it. River's own AI models write nothing for a local chat: no first line or step labels from an AI model, a title from the chat's first words, its history kept in full instead of shortened, no other model taking over a request the local model declines, its content left out of quality monitoring, and the greeting that opens a new chat not reading it. What still leaves your computer: River's cloud stores the chat; River's memory, if your plan includes it, is searched with each request and is sent what you wrote, whichever model you pick (nothing from an Incognito chat is saved to it); the tools River uses for you (web search, pictures, voice, connected apps, River's browser and cloud computer) send what each needs; and Quick ask always runs on cloud models, as does anything it reads to answer you. How Privacy Works lists each one.
- You have the computer for it. River suggests a model for a laptop, a gaming computer or a workstation, and downloads it for you.
- Cost matters. A local model's answers cost nothing beyond the computer you run it on.
You can also mix: use a local model in one chat and a cloud model in another.
Setting it up#
Local models work with River on your computer: the computer that runs the model.
- In a chat, open the model menu and choose Local models.
- Choose Set up. River installs and starts Ollama for you; if it can't on your computer, install Ollama yourself and come back.
- Download one of the models River suggests for your computer, or choose Browse more models to see others.
- Pick the model in the same menu. That chat now runs on it.
Which model to pick#
The Local models menu suggests three, by the kind of computer you have:
- Laptop: quick everyday help (a 4.7 GB download)
- Gaming Computer: smarter, more dependable answers (9 GB)
- Professional: top quality for big jobs (43 GB)
Browse more models lists others. River's work needs a model that can use tools, such as opening a file or searching, and bigger models handle long jobs with many steps better. If a small model gets stuck on a job, try a larger one, or pick a cloud model for that chat.
Which model a chat uses#
You choose the model for each chat in its model menu. A chat on a local model stays on it: River doesn't send some requests to a cloud model, and no other model takes over when the local one can't do something. To use a cloud model, pick one in that chat's menu.
What still needs the cloud#
A local model writes the chat's answers. These still use their own services, whichever model you pick:
- Pictures, video and voice, which the services that make them handle (the Privacy Policy lists them)
- Web search, and the apps you connect, such as Gmail or Stripe
- Everything in the Privacy matters note above: River's cloud, memory, River's browser and cloud computer, and Quick ask
Tips#
- Mac Studio M3 Ultra and NVIDIA RTX 4090+ are the sweet spot for personal use.
- A local chat stays on its model. For a task that's too much for it, pick a cloud model in that chat's menu.