Choosing AI models
Hemmia can use two kinds of models, and you can mix them: pick a model for each chat.
| Local models | Cloud models | |
|---|---|---|
| Where they run | On your hub | At the provider (Anthropic, OpenAI, OpenRouter) |
| Your messages | Stay in the house | Go to the provider for the chats where you pick them |
| Cost | Electricity | You pay the provider for what you use |
| Speed and ability | Good for everyday things, slower on small computers | The strongest models available |
Local models
Local models need Ollama on the hub. The more memory your computer has, the bigger (and better) the models it can run.
Open the models settings
In Hemmia, open Settings → Models. Admins see the section For the whole family.
Pick a model
Under Add a model, Hemmia lists models that fit your computer's memory, with the best fit marked Recommended. Models that need more memory are greyed out.
Download it
Tap Download. A model is a few gigabytes, so this can take a while; you can leave the page, the download carries on.
Once downloaded, the model appears in everyone's model list. Each person can set their own Default model for new chats.
LM Studio and other local servers
If you already run LM Studio, llama.cpp or vLLM on this computer or another one in your home network, add it under Other local servers with its address. Its models then show up in Hemmia too.
Cloud models with your own key
Each person adds their own key; it is stored encrypted with their passphrase, so nobody else, not even the admin, can see or use it.
Get a key
Create an API key with your provider: Anthropic at platform.claude.com, OpenAI at platform.openai.com, or OpenRouter at openrouter.ai. Set a monthly spending limit there too.
Add it in Hemmia
Open Settings → Models → Cloud AI, choose the provider, paste the key and tap Check and save. Hemmia checks the key with the provider without sending any of your chats.
Optionally, set a limit
Under Monthly limit in US dollars, Hemmia can stop using cloud models for you once your estimated spending reaches the limit.