The assistant on your own computer¶

The setup assistant normally borrows a big AI from a company (ChatGPT, Claude, Grok, Gemini). It can also use a local model: a small AI that runs on your own computer, through a free program called Ollama (or LM Studio).

	Big AI (AI app or API key)	Local model
Cost	A subscription or pay per use	Free
Privacy	Your messages go to the AI company	Nothing leaves your computer
Gets it right	Almost always	Often, but not always
Speed	A few seconds	Seconds with a graphics card, minutes without one
Needs	An account	A few GB of disk space; ideally a graphics card

Be honest with yourself about what to expect

A local model is a small brain. It is good at questions ("what does my Raid alert do?") and simple builds ("make a !lurk command"). It gets bigger jobs wrong much more often: it may make the command but forget to connect the action, pick the wrong step, or say "done" when it is not. Always check what it made on the Actions and Commands pages. For big setup jobs, a big AI is the better tool.

Which model to use¶

Cuebot was tested with these. "Memory" is the graphics card memory (VRAM) the model needs.

Your computer	Model	Download	Memory	How it did in Cuebot's tests
Graphics card with 8 GB or more	Qwen3.5 9B (`qwen3.5:9b`)	6.6 GB	about 6.2 GB	Right 14 times out of 15
Graphics card with 4 to 6 GB, or no graphics card	Qwen3.5 4B (`qwen3.5:4b`)	3.3 GB	about 3.8 GB	Right 8 times out of 15: fine for questions and simple builds, often wrong on bigger ones

Other models also work if they can "call tools" (Ollama marks them with tools). Gemma 4 E4B came third in the tests. Very small models (3B and below) were not reliable enough to recommend.

No graphics card? It works, but slowly: a simple build took about half a minute to several minutes on a fast desktop processor, and longer on a laptop. Use the 4B model, and be patient.

Setting it up¶

Install Ollama from ollama.com (Windows, Linux and Mac). It starts by itself and runs in the background.
Download a model. Open a terminal (on Windows: PowerShell) and type one of these, then wait for it to finish:
```
ollama pull qwen3.5:9b
```
```
ollama pull qwen3.5:4b
```
In Cuebot, go to Home > Assistant and pick Local model (Ollama, LM Studio) in AI. The dot turns green when Cuebot finds Ollama and a model.
Leave Model empty to let Cuebot pick (it prefers Qwen3.5), or pick one from the list.

Using LM Studio instead? Start its server (Developer tab, or lms server start), load a model, set its context length to at least 16384, and type http://127.0.0.1:1234 in Server.

The options¶

Server: where Ollama or LM Studio listens. Leave it empty for Ollama on this computer.
Memory (tokens): how much of the conversation the model can keep in mind (its "context"). 32768 is right for most people. Lower it only if your graphics card runs out of memory; below about 16000, Cuebot's tools no longer fit.
Let it think first: off by default. Thinking made the models much slower (a simple build went from about half a minute to five minutes without a graphics card) and did not make them more accurate in Cuebot's tests.
Give it every tool: off by default. Small models get Cuebot's tools for actions, commands and timers, plus a way to find the others (overlays, polls, points and so on) when they need them. With all of Cuebot's tools (over a hundred) at once, small models got confused and failed about twice as often. Tick it only for big models (30B and up).

Changes apply from the next New chat.

Problems you may see¶

What happens	Why	What to do
It says "done" but the command does nothing	It made the command but did not connect an action to it, or used the wrong trigger	Check the action's Triggers; tell it what is missing ("the !lurk command has no action, add one")
It asks you to type "yes" before deleting	Small models sometimes ask in words instead of using the Approve button	Answer "yes"; Cuebot still shows Approve before anything is deleted
It changes something you did not ask for	It misread the request	Click Undo these changes under its answer, and ask again more simply. Before a big job, a backup (Settings > Backups) is a good idea
An empty answer, or "the model server returned an error"	Small models sometimes stop or write a broken request. Cuebot already retries once and nudges once	Send the message again, or press New chat
"The conversation has filled this model's memory"	Long conversations do not fit	Press New chat, or raise Memory if your graphics card has room
Very slow answers	No graphics card, or the model does not fit in its memory	Use the 4B model, close games and other heavy programs, keep Let it think first off
"Ollama is not running"	Ollama is closed	Start Ollama (on Windows, from the Start menu) and press Refresh
"Ollama is running but has no models yet"	No model downloaded	Run `ollama pull qwen3.5:4b` and press Refresh

Tips for better results

Ask for one thing at a time: "make a !pausealerts command for moderators", then "make it pause the alerts and the speech".
Use the exact names you see in Cuebot ("the action called Raid alert").
Start a New chat for each new job, so the model's memory stays clear.

How Cuebot tested these models¶

Cuebot sent typical requests to each model through the real assistant, five times each, and checked what was actually built instead of trusting the answer (October 2026, Ollama 0.40.13, RTX 4090):

Request	Qwen3.5 9B	Qwen3.5 4B
"Make a !lurk command that thanks the viewer by name"	5 of 5	3 of 5
"Moderator-only !pausealerts and !resumealerts that pause and resume alerts and text-to-speech"	4 of 5	1 of 5
"Every 15 minutes, post my socials in chat"	5 of 5	4 of 5

Each took a few seconds with a graphics card. Questions about your setup ("which actions use a chat command?") and deleting something (with Approve) also worked in Cuebot's tests. New models come out often; if you try another one, check its results the same way. The tests are part of Cuebot's source code (docs/ASSISTANT.md explains how to run them).