The AI models

Homefellow uses two AI models, which can be the same one:

• the reading model reads each new thing the family shares (an answer, a note, a calendar change) and updates what Homefellow knows;
• the questions model chooses the next question or lookup.

Both run on the household’s own computers. Homefellow talks to them through goose, which is bundled with the server and checked by its checksum.

What works

• A model server with an OpenAI-compatible API: Ollama, LM Studio, llama.cpp’s server or vLLM, on the household’s computer or another one at home.
• A capable open model. Homefellow is developed and tuned on GLM. In our testing so far, larger models read more reliably and small ones tend to miss things.
• A powerful computer made for AI, or a large graphics card. Homefellow is aiming to run on smaller computers as models get better.

The exact requirements (which models, how much memory and how fast) will be published with the first release, from what we’ve tested.

How setup finds them

homefellow models discover (the setup page’s “AI models” step) looks for model servers on the usual ports:

Server	Port	Asks
Ollama	11434	/api/tags
LM Studio	1234	/v1/models
llama.cpp	8080	/v1/models
vLLM	8000	/v1/models

It looks on this computer, on any other computer you name (--host 192.168.1.40) and at any address you type (--url http://gpu-box:8000/v1), giving each 1.5 seconds. It leaves out embedding and speech models, then recommends a pair, with a verdict:

• good: 30 billion parameters or more, or the GLM Homefellow is tuned on;
• maybe: 14 to 30 billion;
• too small: under 14 billion;
• unknown: its size couldn’t be read;
• none: no model server answered.

Choosing

sudo homefellow models choose --recommended
sudo homefellow models choose --url http://127.0.0.1:11434 --model qwen3:32b
sudo homefellow models choose --url URL --model NAME --explorer-url URL2 --explorer-model NAME2

A new choice takes effect at once, with no restart. The choice is written as profiles in the data folder’s profiles/, with the settings pointing at them.

The test

sudo homefellow models test

It reads an invented note about a dog’s rabies booster and a passport into a throwaway household, then chooses one question. It makes real model calls, so it takes as long as one reading and one question: often a minute or two.

Verdict	Means
good	both things were found and a question was chosen in good time
slow	it works, but slowly
too small	it answered, but missed things: try a larger model
failed	a model didn’t answer (“Failed: the reading model didn’t answer”), with the address it tried: check its server is running there and the model is loaded. A model that answered but never with something Homefellow could use is failed too. The words say which

The result is kept. The doctor shows it.