Shows the state of the optional local AI model: switched on or off, ready or not, the chosen model, the downloaded
runtime, this PC's hardware and which model size JoyPro recommends for it. engine llm recommend prints only the
recommendation. Settings > Local AI shows this status (Refresh reads it again), and every AI action checks it first.
engine llm status --text
Advanced
Without options the hardware is detected (RAM, graphics cards with video memory, cores, AVX2).
--ram-gb, --vram-gb, --gpu <name>, --gpu-vendor nvidia|amd|intel, --cores, --avx2 true|false replace the
detected values, to see what JoyPro would recommend for another PC.
The recommendation: about 8 GB video memory: the large model on the graphics card; about 4 GB: the medium one; else
on the processor by memory and AVX2 (medium, small, tiny), or none below about 4 GB of memory.
Output: JSON (Enabled, Ready, SelectedModel, Hardware, Recommendation, the runtime, your own llama.cpp, the
models with size, licence, installed, recommended), or --text. recommend: the recommendation alone.
Exit code 1 with Invalid --ram-gb value: ... (and the same for the other values) or an unknown argument.
Files: %LOCALAPPDATA%\JoyPro2\Llm\settings.json, the model list DB\Llm\models.json.
What it does The hardware, the model JoyPro recommends for it (size, quantization, GPU or CPU), downloaded models and runtime; the options replace detected values (what-if).
Downloads a local AI model and the llama.cpp runtime it runs in, checks both, and switches the model on. Nothing
downloads without --accept-license. Settings > Local AI > Download and switch on (after ticking I accept the
size and the licenses above) runs it.
engine llm download --accept-license --text
Advanced
<model> (or --model <id>): a model id of DB\Llm\models.json; without it, the chosen or recommended model.
--accept-license: without it the command only names the download (size, source, licence) and fails.
--runtime auto|cpu|vulkan|cuda: the llama.cpp build (auto: Vulkan with a graphics card, else CPU); an explicit
choice is remembered. --no-runtime downloads the model only.
The model comes from its pinned source and is checked by SHA-256; a model already downloaded and verified is not
fetched again. A stopped download keeps its part and the next download resumes it.
After success the model is chosen and the local AI is switched on. No profile is touched.
Technical internals
Files: %LOCALAPPDATA%\JoyPro2\Llm\models\<file> with <file>.sha256, the runtime in Llm\runtime, the progress in
Llm\download-progress.json (the GUI follows it), Llm\settings.json.
Output: JSON (Model, Runtime, Consent, Messages, Path), or --text.
Exit code 1 with Downloading needs your consent. ... Run again with --accept-license., Unknown model '...',
Unknown runtime '...', Download failed: ... or the cancel message.
What it does Downloads a model (positional or --model; default: the chosen model, else the recommended one) and llama.cpp's server, checks their SHA-256 and switches the model on; without --accept-license only the size and licenses are shown.
Option
Values
Meaning
--model
<model>
The model id to download (same as the positional <model>); default: the chosen model, else the recommended one.
--accept-license
switch (no value)
--runtime
autocpuvulkancuda
--no-runtime
switch (no value)
Every value as a command line
engine llm download --runtime auto
engine llm download --runtime cpu
engine llm download --runtime vulkan
engine llm download --runtime cuda
Enable switches the local model on, with a model also chooses it; disable switches it off and keeps the chosen model; for for relations from-text, the assistant, relations chat, relations duplicates --instruction, arrange ask and relations plan; local = your own llama.cpp / GGUF from llm use-local
For beginners
Switches the optional local AI model on, optionally choosing which downloaded model to use; engine llm disable
switches it off. Settings > Local AI > Use the local AI model runs it.
<model>: a model id of DB\Llm\models.json; without it the chosen model stays. local: your own llama.cpp and
GGUF set with engine llm use-local (refused when none is set). llm disable takes no model: it switches the
model off and keeps the chosen one.
It only changes the setting; nothing is downloaded or started. Whether the model is usable (downloaded, runtime
there) shows engine llm.
Output: the settings as JSON, or --text (Local AI model enabled (<model>). / ... disabled ...).
Exit code 1 with No own llama.cpp / GGUF set (engine llm use-local --server <llama-server.exe> --model <file.gguf>)., Unknown model '...'. Models: ... or llm disable takes no model: ....
What it does Switches the local AI model on (optionally choosing a downloaded model, or local = your own llama.cpp and GGUF) or off (llm disable, which takes no model and keeps the chosen one). It is used by relations from-text, the assistant, relations chat, relations duplicates --instruction, arrange ask and relations plan.
Looks for a llama.cpp server (llama-server.exe) and GGUF model files that are already on this PC, and names the pair
that suits the hardware. It only reads. Settings > Local AI > Find on this PC runs it and fills in the paths.
engine llm find-local --text
Advanced
Searched by default: llama-server.exe on the PATH, folders like llama.cpp, AI, LLM, models and tools
on the fixed drives, your Downloads, the Hugging Face cache, LM Studio's model folders and JoyPro's own model
folder. --root <folder> (repeatable) searches only these folders.
At most 7 folder levels deep and about 25 seconds; Ollama's own server and blobs are skipped.
Only language models count (no vision projectors, vocab-only or embedding files, from 50 MB; of a split model the
first part).
Recommended: the largest model that fits completely in the graphics card's memory, else the smallest that fits the
memory; a GPU build of the server when there is a graphics card.
Makes JoyPro use a llama.cpp server and GGUF model you already have instead of a download, and switches the local AI
on. Nothing is downloaded. Settings > Local AI > Use these files runs it with the two paths; Stop using them
runs --clear.
--server: llama-server.exe (or its folder); --model: the .gguf file. A missing one is taken from the search
of engine llm find-local.
--gpu-layers auto|all|<n> (auto: llama.cpp decides), --context <n> (512 or more), --template server|chatml
(server: the GGUF's own chat template).
--clear forgets the two paths; when they were in use the local AI is switched off. Your files stay where they are.
Technical internals
Output: JSON or --text (Model: local, the runtime path and the message Using your llama.cpp (...) with ...; the local AI model is on. Nothing was downloaded.).
Exit code 1 with llama-server.exe not found: ..., GGUF model not found: ..., No llama-server.exe found (searched ...), Invalid --gpu-layers value: ..., Invalid --context value: ... (512 or more). or Invalid --template value: ....
What it does Uses your own llama.cpp server and GGUF model instead of a download (without paths: what find-local recommends) and switches the model on; --clear forgets them.
Option
Values
Meaning
--server
<llama-server.exe>
--model
<file.gguf>
--gpu-layers
autoallN
--context
switch (no value)
--template
serverchatml
--clear
switch (no value)
Every value as a command line
engine llm use-local --gpu-layers auto
engine llm use-local --gpu-layers all
engine llm use-local --gpu-layers N
engine llm use-local --template server
engine llm use-local --template chatml
Loads the chosen model now; started from JoyPro it stays loaded until the idle stop
For beginners
Loads the chosen local AI model now, so the next AI request answers without the loading wait. Settings > Local AI >
Start now runs it. Started from JoyPro, the model stays loaded until the idle stop; from a command line it ends
again when the command ends.
engine llm start --text
Advanced
Needs a usable model: the local AI switched on and a downloaded model (with its runtime) or your own llama.cpp and
GGUF.
The model runs in its own llama-server process on this PC. engine llm server shows it, engine llm stop ends it
and frees its memory.
No profile is touched.
Technical internals
Output: JSON (Started, Model, ModelId, Pid, LoadMs, StaysLoaded), or --text (Started the local AI model <name> (process <pid>, <s> s); ...).
Exit code 1 with The local AI model cannot start: <reason> (for example the model is off or not downloaded) or
The local AI model (<name>) did not start: ....
What it does Loads the chosen model now; started from JoyPro it stays loaded until the idle stop.
Tells whether JoyPro's local AI model is loaded right now, which model and how much memory it uses.
engine llm stop ends it and frees that memory.
engine llm server --text
Advanced
Started from JoyPro, the model stays loaded between requests until the idle stop; from a command line every request
starts and ends its own model process, so server usually says Not running. there.
stop ends only the process JoyPro recorded as its model server (not another program's llama.cpp) and forgets it.
No profile is touched.
Technical internals
Output of server: JSON (Running, and when running Pid, Model, LastUsedUtc, IdleMinutes, Busy,
WorkingSetBytes), or --text (Running: <model> (process <pid>) / Not running.). Output of stop: JSON
(Stopped, Pid), or --text (Stopped the local AI model (process <pid>). / The local AI model is not running.). Exit code 0; 1 with an extra argument.
File: %LOCALAPPDATA%\JoyPro2\Llm\server.json (the running server; removed by stop).
A test helper, not for players: a small stand-in for the llama.cpp server that JoyPro's automatic tests use, so the
AI paths and the model download can be tested without a real model or the internet. It listens on this PC only and
ends by itself.
--port N (default 0 = a free port); --port-file <file> receives the port it got.
--response <file>: the answer to every completion request; answers separated by a line #next are given one per
request, the last one repeated. Without it the answer is {}.
--serve <folder>: files of that folder are downloadable (with resume), for the downloader tests.
--log <file>: every request is logged. --lifetime <s> (default 300): it ends after that many seconds at the
latest.
Only the loopback address; nothing of the profile or settings is touched.
Technical internals
Answers GET /health, POST /completion, POST /v1/chat/completions and GET /files/<name>.
Exit code 0 when the lifetime is over; 1 with Invalid --port value: ..., Invalid --lifetime value: ... or an
unknown argument.
The engine starts it itself instead of llama.cpp when the variable JOYPRO_LLM_FAKE_RESPONSE names a response file.
What it does A test stub that answers like llama-server and serves files with HTTP Range (no model involved).