Settings and models

Change settings and choose a model.

Settings

pionus works without configuration. Settings change the style, the device or the model.

pionus config list --show-origin
# everything, and the layer each value comes from
pionus config set model.device cpu
# global
pionus config set --repo style.convention conventional
# .pionus.toml, shared through git
pionus config get style.convention
pionus config unset model.device
pionus config init
# a file with every key commented out

A value comes from the strongest layer that sets it:

command-line flags (--backend, --device, --model-dir);
environment variables (PIONUS_*);
the repository's .pionus.toml, at the root of the work tree;
the global ~/.config/pionus/config.toml;
the built-in defaults.

Common tasks

To	Run
enforce Conventional Commits for a team	`pionus config set --repo style.convention conventional`
get faster NPU replies that see less of the diff	`pionus config set model.max_prompt_tokens 512`
use another GPU on a multi-GPU machine	`pionus config set model.vulkan_device 1`
keep replies short	`pionus config set model.max_new_tokens 64`

Every key is in the configuration guide.

Keeping the model loaded

On vulkan and cpu the model stays loaded for 10 idle minutes. Change it with model.keep_loaded; 0 turns it off. See the measurements for what it saves.

Asking for the type

With pionus config set model.ask_type low_confidence, --commit asks for the type when the model is unsure.

Off by default. It works for Conventional Commits, on vulkan and cpu, in a terminal. See the measurements.

The update notice

pionus makes no request by itself. To hear about new releases:

pionus config set update.check true

Once a day at most, it asks GitHub for the latest release. GitHub sees your IP address and pionus/<version>, nothing else. The hook never checks. PIONUS_UPDATE_CHECK=false turns it off for one command.

Models

pionus ships no model. model use downloads one from Hugging Face and checks it against a SHA-256. The project's own fine-tunes are on its Hugging Face account.

pionus model
# a picker, in a terminal
pionus model list
# the recommended models and those from disk, and which is in use
pionus model use qwen2.5-coder-7b
# download it from Hugging Face and use it
pionus model use ~/models/my.gguf
# a GGUF file from disk, unverified, remembered

Model	Size	Status	Devices
`qwen2.5-coder-1.5b`	1.1 GB	validated	all
`qwen2.5-coder-0.5b`	491 MB	experimental	vulkan, cpu
`qwen2.5-coder-7b`	4.7 GB	experimental	vulkan, cpu
`pionus-1.5b-v0.1`	986 MB	experimental	vulkan, cpu
`pionus-0.5b-v0.1`	397 MB	experimental	vulkan, cpu
`pionus-0.5b-v0.2`	397 MB	experimental	vulkan, cpu
`pionus-1.5b-v0.2`	986 MB	experimental	vulkan, cpu
`pionus-0.5b-v0.3`	397 MB	experimental	vulkan, cpu
`pionus-1.5b-v0.3`	986 MB	experimental	vulkan, cpu

The default is the only model the NPU and the Intel GPU can run.
Experimental models are not fully tested.
pionus-* models are fine-tuned for this prompt.
After switching, run pionus bench again.

Times per model are in the measurements. Licences are in the models guide.