Models: pick and manage the AI model
For anyone choosing what writes their replies, posts and chats. Every model on this page runs on this computer, and one choice covers the side panel and Chat alike.
Open it from Models in Riff Home's sidebar, by right-clicking the Riff icon in the toolbar and choosing Models, or by clicking the dot on the Riff logo in the side panel once the model is ready Riff Home: what it is and how to open it.
Riff writes with
The card at the top names the model in use, like Riff writes with Gemini Nano, with Replies, posts and Chat all use it. Under it are four choices:
| Choice | Its line | What it does |
|---|---|---|
| Auto | Gemini Nano when Chrome has it, else Gemma 4 | Riff picks for you Which AI model? Gemini Nano or Gemma 4 |
| Gemini Nano | Built into Chrome | Chrome's own model. Nothing to download from Riff |
| Gemma 4 | The size you use, like Gemma 4 E2B | Riff's own download. Needs a size downloaded first |
| Model app | LM Studio or Ollama | A model running in LM Studio or Ollama on this computer. Greyed out until you connect one |
Built into this browser
Gemini Nano shows Chrome's status for it: Ready, Not downloaded with a Download button, Downloading, Can't run here or Not in this browser. It writes 5 languages The languages Riff writes in.
Gemma 4, downloaded by Riff
Gemma 4 is Google's open model. Riff downloads it once and runs it on your graphics chip, and it writes 16 languages. There are two sizes:
| Size | Its line | Download |
|---|---|---|
| Gemma 4 E2B, the Default size | Fast. Good for replies, posts and everyday chat. | About 3 GB |
| Gemma 4 E4B | Better writing and answers. Needs a strong GPU. | About 5 GB |
Before you download, each size says how it will run here, from your graphics chip and memory: Runs well here, Slow here, or Can't run here (no WebGPU). On a laptop without a strong graphics chip, stay with E2B.
- Click Download next to a size. A bar fills; keep the tab open until it's done Download Gemma 4.
- Click Use. The message says Riff now writes with Gemma 4 E2B. It loads onto your GPU on the first riff. The row then reads In use.
- To free the space, click Delete, then Delete?. Only one size is loaded at a time.
At the top right, Models on disk adds up what Riff has downloaded, and the line under it says what the browser reports about your graphics chip, like Graphics: WebGPU ready · fast 16-bit math.
LM Studio
Already run bigger models in LM Studio? Riff can write with them, still on this computer.
- In LM Studio, load a chat model and start the local server: Developer › Start server. Turn on its Enable CORS option.
- In Riff Home › Models, under Model apps on this computer, find LM Studio (
localhost:1234) and click Connect. - The first time, the browser asks to let Riff reach apps on this computer. Allow it: it covers
localhostand127.0.0.1only. - The card reads Connected. Pick a model in its menu and click Use this model. The message says which, like Riff now writes with qwen3-14b in LM Studio.
Ollama
- Ollama refuses browser extensions unless you allow them. Quit Ollama, then start it from a terminal with
OLLAMA_ORIGINS="chrome-extension://*" ollama serve. On a Mac with the Ollama app, runlaunchctl setenv OLLAMA_ORIGINS "chrome-extension://*"in Terminal instead, then restart the app. - Pull a chat model in Ollama if you haven't yet.
- In Riff Home › Models, click Connect on Ollama (
localhost:11434), allow the browser's question, pick a model and click Use this model.
If the card reads Needs one setting, Ollama is running but still refuses Riff: it shows the exact command for your browser LM Studio or Ollama won't connect.
How model apps work with Riff
- Riff talks to
localhostonly, and to neither app until you click Connect. It never reaches a server on another computer, even if you point it at one. - Embedding models aren't offered, only chat models. A model's hidden reasoning isn't shown in drafts.
- Replies, posts and Chat all use the model you picked. If the app is closed later, the card reads Not found and Riff can't write until you start it again or pick another model.