Models and memory

TuxAide runs a model locally through Ollama. The installer (and tuxaide setup) picks the one that fits your RAM:

RAM	Model	Download	Answers
8 GB or more	qwen2.5-coder:7b	4.7 GB	Best
5–8 GB	qwen2.5:3b	1.9 GB	Good
3–5 GB	qwen2.5:1.5b	986 MB	Simpler, but usable
under 3 GB	qwen2.5:0.5b, or a remote backend	398 MB	Basic

Download sizes are from the Ollama registry. In a container or VM with a memory limit, the limit counts, not the host's RAM. Smart RAG needs 5 GB or more.

Switch models

tuxaide model                 # the models in Ollama, the current one marked
tuxaide model llama3.2        # switch; offers `ollama pull` if it isn't downloaded

Any model in the Ollama library works. Coding-tuned models give the best commands.

Speed

• The first words appear as soon as the model writes them — answers stream in.
• Pre-loading: when a terminal opens, TuxAide asks Ollama to load the model (prewarm: once per 10 minutes by default, off on machines under 8 GB, or always). keep_alive (10 min) is how long Ollama keeps it in memory after a question.
• Repeated questions come from the cache instantly.
• A GPU makes answers much faster but isn't needed.

tuxaide --timing shows how long recent questions took (from ~/.config/tuxaide/logs/perf.log). tuxaide doctor warns when free memory is too low for the model and suggests a smaller one.