Command reference
Every yonga command and subcommand, with its options and defaults as yonga <command> --help prints them in yonga 0.6.0, plus one example each. The guides explain when to use each command; this page is for looking things up. The full behaviour of every command, including its --json document, is specified in CLI_SPEC.md.
In the option tables, a dash in the Default column means --help shows no default: the option is off, or yonga decides from the artifact, the registry or the configuration, as the description says. Every command and command group also accepts --help, which is not repeated below.
Global options
yonga [OPTIONS] COMMAND [ARGS]...
Local AI model manager for Intel AI PCs (OpenVINO, Optimum Intel, Hugging Face).
| Option |
Value |
Default |
Description |
--config |
PATH |
~/.config/yonga/config.toml |
Configuration file to load instead of the default. |
--data-dir |
PATH |
~/.local/share/yonga |
Where prepared artifacts and logs are kept. |
--cache-dir |
PATH |
~/.cache/yonga |
Where the managed temp and compile cache live. |
--offline |
|
off |
Never contact the Hugging Face Hub. |
--json |
|
off |
Emit one JSON document on stdout and no decoration. |
--verbose, -v |
|
off |
Explain what is happening. |
--debug |
|
off |
Include raw error text in failures. |
--version, -V |
|
|
Show the version and exit. |
Global options go before the command name:
yonga --offline --debug prepare Qwen/Qwen3-0.6B
--json is the one global option that commands also accept after their own name (yonga list --json). chat and serve refuse it with exit 70. The default directories follow XDG_CONFIG_HOME, XDG_DATA_HOME and XDG_CACHE_HOME when those are set; yonga registry paths prints the ones in use. Settings are taken from the command line first, then YONGA_* environment variables, then the configuration file.
yonga with no command prints the command list and exits 70, as does any command line that cannot be parsed (an unknown option, a missing argument, a value of the wrong type). --help exits 0. See Exit codes for every code.
Commands at a glance
| Command |
Purpose |
doctor |
Check this host: OpenVINO, devices, NPU access, storage and authentication. |
search |
Search the Hugging Face Hub and show what the registry knows about each result. |
inspect |
Describe a model and what preparing it on this host would involve. |
prepare |
Download, convert, validate and probe a model into a prepared artifact. |
list |
List the artifacts prepared on this host. |
remove |
Remove a prepared artifact and its compile-cache entries. |
probe |
Compile a prepared artifact on a device and report what worked. |
chat |
Chat with a prepared artifact, locally. |
serve |
Serve a prepared artifact through an OpenAI-compatible HTTP API. |
eval |
Measure how far CANDIDATE's predictions move from REFERENCE's (KL, top-1, perplexity). |
publish |
Upload a prepared artifact to the Hugging Face Hub with a generated model card. |
bench |
Benchmark a prepared artifact and save a reproducible JSON result. |
logs |
Show the job logs kept on disk, most recent first. |
registry |
Inspect, validate and update the compatibility registry. |
env |
List and remove the managed environments exports run in. |
cache |
Show and prune the OpenVINO compile cache. |
self |
Put the yonga AppImage on PATH as 'yonga', or take it off again. |
connect |
Register a prepared artifact with a client that talks to 'yonga serve'. |
Where a command takes MODEL_OR_ARTIFACT or ARTIFACT, you can pass an artifact id from yonga list (or a unique prefix of at least six characters), the artifact's directory, or a Hugging Face repository id. For a repository id with several prepared artifacts, the read-only commands use the newest ready one; remove refuses and lists them.
yonga doctor
Check this host: OpenVINO, devices, NPU access, storage and authentication.
| Option |
Value |
Default |
Description |
--full |
|
off |
Also report the loaded registry. |
--json |
|
off |
Emit the report as JSON. |
Exits 0 when a runtime and device are usable and nothing needs attention, 2 when usable with warnings, 10 when no OpenVINO runtime or device is usable. Maintenance explains each section.
yonga search
yonga search [OPTIONS] QUERY
Search the Hugging Face Hub and show what the registry knows about each result.
| Argument |
Description |
QUERY |
Words to search the Hub for. Required. |
| Option |
Value |
Default |
Description |
--limit |
integer, 1–100 |
20 |
How many results to show. |
--device |
text |
– |
Show compatibility for one device: npu, gpu, cpu or auto. |
--json |
|
off |
Emit the results as JSON. |
yonga search qwen3 --device npu --limit 10
yonga inspect
yonga inspect [OPTIONS] MODEL
Describe a model and what preparing it on this host would involve. Downloads no weights.
| Argument |
Description |
MODEL |
Hugging Face repository id. Required. |
| Option |
Value |
Default |
Description |
--revision |
text |
– |
Branch, tag or commit sha. |
--files |
|
off |
Also list the repository files. |
--json |
|
off |
Emit the report as JSON. |
yonga inspect Qwen/Qwen3-0.6B
yonga prepare
yonga prepare [OPTIONS] MODEL
Download, convert, validate and probe a model into a prepared artifact. See Preparing models.
| Argument |
Description |
MODEL |
Hugging Face repository id. Required. |
| Option |
Value |
Default |
Description |
--revision |
text |
– |
Branch, tag or commit sha. |
--device |
text |
auto |
auto, npu, gpu or cpu. |
--profile |
text |
– |
Conversion profile id (yonga registry profiles lists them). Without it the registry chooses. |
--source |
text |
– |
Convert this repository instead of the requested one. |
--trust-remote-code |
|
off |
Allow the conversion to execute Python published in the repository. |
--ignore-exporter-requirements |
|
off |
Convert even though a declared exporter package version is unmet, or the registry knows this export fails with the installed versions. Deliberate and separate: --yes never enables it. |
--no-managed-env |
|
off |
Never build a managed export environment; refuse instead when the installed packages cannot meet the exporter requirements. |
--gguf-file |
text |
– |
For a GGUF repository, the file to import. By default yonga picks the first preferred quantization whose tensor types OpenVINO GenAI can read. |
--fresh-export |
|
off |
Export again even when a failed job left a completed export with identical inputs that could be reused. |
--yes, -y |
|
off |
Answer ordinary questions with yes. |
--force |
|
off |
Re-prepare even if an artifact exists. |
--no-repair |
|
off |
Never modify the exported IR. |
--json |
|
off |
Emit the outcome as JSON. |
yonga prepare Qwen/Qwen3-0.6B --device npu
yonga list
List the artifacts prepared on this host.
| Option |
Value |
Default |
Description |
--clean |
|
off |
Retire stale staging and temp directories under the retention policy. |
--json |
|
off |
Emit the catalogue as JSON. |
yonga remove
yonga remove [OPTIONS] ARTIFACT
Remove a prepared artifact and its compile-cache entries. The Hugging Face download cache is never touched.
| Argument |
Description |
ARTIFACT |
Artifact id, repo id or path. Must match exactly one artifact. Required. |
| Option |
Value |
Default |
Description |
--yes, -y |
|
off |
Do not ask for confirmation. |
--json |
|
off |
Emit the outcome as JSON. |
yonga remove cc24636a2f49356a
yonga probe
yonga probe [OPTIONS] MODEL_OR_ARTIFACT
Compile a prepared artifact on a device and report what worked.
| Argument |
Description |
MODEL_OR_ARTIFACT |
Artifact id or path. Required. |
| Option |
Value |
Default |
Description |
--device |
text |
auto |
auto, npu, gpu or cpu. |
--all-submodels |
|
off |
Compile every submodel, not only the language model. |
--quick |
|
off |
Never build the GenAI pipeline; a deferred raw compile is reported inconclusive. |
--timeout |
SECONDS |
– |
Bound the GenAI pipeline build that settles a deferred compile: run it in a child process and stop it after SECONDS (the device is then inconclusive). Cannot be combined with --quick. |
--json |
|
off |
Emit the report as JSON. |
Exits 0 when any device passed, 2 when none passed but none failed outright either (left inconclusive by --quick or --timeout), otherwise the code of the first failure.
yonga probe cc24636a2f49356a --device npu
yonga chat
yonga chat [OPTIONS] MODEL_OR_ARTIFACT
Chat with a prepared artifact, locally. See Chat, probe, bench and eval for the in-chat commands such as /reset and /exit.
| Argument |
Description |
MODEL_OR_ARTIFACT |
Artifact id or path. Required. |
| Option |
Value |
Default |
Description |
--device |
text |
– |
auto, npu, gpu or cpu. Without it, the default_device setting when it names a device, otherwise the artifact's recommended device. |
--max-new-tokens |
integer |
– |
Output token limit. |
--temperature |
number |
– |
Sampling temperature. |
--top-p |
number |
– |
Nucleus sampling probability. |
--top-k |
integer |
– |
Top-k sampling cutoff. |
--show-reasoning |
|
off |
Display hidden reasoning blocks. |
--system |
text |
– |
System prompt for the session. |
--json |
|
|
Not supported: a conversation is not one document (exit 70). |
yonga chat cc24636a2f49356a
yonga serve
yonga serve [OPTIONS] MODEL_OR_ARTIFACT
Serve a prepared artifact through an OpenAI-compatible HTTP API. Needs the serve extra in a pip install; the AppImage includes it. See OpenAI-compatible server.
| Argument |
Description |
MODEL_OR_ARTIFACT |
Artifact id or path. Required. |
| Option |
Value |
Default |
Description |
--host |
text |
127.0.0.1 |
Address to bind. Loopback unless you say otherwise. |
--port |
integer, 1–65535 |
8000 |
TCP port to listen on. |
--device |
text |
– |
auto, npu, gpu or cpu. Without it, the default_device setting when it names a device, otherwise the artifact's recommended device. |
--max-new-tokens |
integer |
– |
Default output token limit when a request sets none. |
--show-reasoning |
|
off |
Return reasoning blocks instead of hiding them. |
--max-body-mb |
integer, 1–64 |
8 |
Largest request body in MiB; embedded images count towards it. |
--carry-images-forward |
|
off |
Move images sent with earlier messages onto the last one instead of refusing them. |
--json |
|
|
Not supported: a server is not one document (exit 70). |
yonga serve cc24636a2f49356a --port 8000
yonga eval
yonga eval [OPTIONS] CANDIDATE
Measure how far CANDIDATE's predictions move from REFERENCE's (KL, top-1, perplexity). Both must be artifacts of the same weights; the reference is usually a higher-precision one, such as one prepared with --profile cpu-int8-asym-ref.
| Argument |
Description |
CANDIDATE |
Artifact to evaluate. Required. |
| Option |
Value |
Default |
Description |
--reference |
text |
– |
Higher-precision artifact of the same weights to compare to. Required. |
--device |
text |
cpu |
cpu (full precision) or gpu. |
--corpus |
path |
– |
Text to read instead of the shipped corpus (.json or .txt). |
--allow-different-sources |
|
off |
Compare artifacts converted from different repositories (e.g. GGUF vs original). |
--json |
|
off |
Emit the result as JSON. |
yonga eval CANDIDATE_ID --reference REFERENCE_ID
yonga publish
yonga publish [OPTIONS] ARTIFACT
Upload a prepared artifact to the Hugging Face Hub with a generated model card. The repository is private unless you pass --public.
| Argument |
Description |
ARTIFACT |
Prepared artifact id or path. Required. |
| Option |
Value |
Default |
Description |
--repo |
text |
– |
Destination NAMESPACE/NAME on the Hub. Always explicit. Required. |
--public |
|
off |
Create a public repository (the default is private). |
--update |
|
off |
Write into a repository that already exists. |
--license-reviewed |
|
off |
State that you read a license yonga cannot classify and it permits redistributing a quantized derivative. --yes never implies it. |
--dry-run |
|
off |
Check everything and write the card locally; upload nothing. |
--output |
path |
– |
Where to write the card, manifest and upload list. With --dry-run and no --output: ./yonga-publish-<artifact>. |
--yes, -y |
|
off |
Approve the upload without a prompt. |
--json |
|
off |
Emit the outcome as JSON. |
yonga publish cc24636a2f49356a --repo NAMESPACE/Qwen3-0.6B-openvino-npu --dry-run
yonga bench
yonga bench [OPTIONS] MODEL_OR_ARTIFACT
Benchmark a prepared artifact and save a reproducible JSON result.
| Argument |
Description |
MODEL_OR_ARTIFACT |
Artifact id or path. Required. |
| Option |
Value |
Default |
Description |
--device |
text |
– |
auto, npu, gpu or cpu. Without it, the artifact's recommended device. |
--suite |
text |
standard |
quick, standard or long. |
--runs |
integer |
– |
Measured runs per prompt. Without it, the bench_runs setting (3). |
--warmups |
integer |
– |
Warmup runs per prompt. Without it, the bench_warmups setting (1). |
--json |
|
off |
Emit the result as JSON. |
yonga bench cc24636a2f49356a --suite quick
yonga logs
yonga logs [OPTIONS] [JOB_ID]
Show the job logs kept on disk, most recent first. Without JOB_ID, lists the 20 most recent jobs.
| Argument |
Description |
JOB_ID |
Job id, or a prefix. Optional. |
| Option |
Value |
Default |
Description |
--raw |
|
off |
Print the JSON Lines file verbatim. |
--bundle |
|
off |
Write a redacted, anonymised .tar.gz for a bug report (log, state, manifest, host report). |
--output |
file |
./yonga-report-<job>.tar.gz |
Where to write the --bundle archive. |
--json |
|
off |
Emit the records as JSON. |
yonga logs de862650 --bundle
yonga registry
yonga registry [OPTIONS] COMMAND [ARGS]...
Inspect, validate and update the compatibility registry. See Maintenance.
registry validate
yonga registry validate [OPTIONS]
Load every registry layer and check it is internally consistent. Exits 0 when valid, 2 when an installed remote bundle was ignored, 70 when problems are found.
| Option |
Value |
Default |
Description |
--json |
|
off |
Emit the result as JSON. |
registry profiles
yonga registry profiles [OPTIONS]
List the conversion profiles --profile accepts.
| Option |
Value |
Default |
Description |
--device |
text |
– |
Only profiles for this device. |
--json |
|
off |
Emit the profiles as JSON. |
yonga registry profiles --device npu
registry explain
yonga registry explain [OPTIONS] MODEL
Show which registry rules matched a model, and why they won.
| Argument |
Description |
MODEL |
Hugging Face repository id. Required. |
| Option |
Value |
Default |
Description |
--device |
text |
– |
Explain for this device only. |
--json |
|
off |
Emit the explanation as JSON. |
yonga registry explain Qwen/Qwen3-0.6B --device npu
registry paths
yonga registry paths [OPTIONS]
Show where registry layers, configuration and managed data live.
| Option |
Value |
Default |
Description |
--json |
|
off |
Emit the paths as JSON. |
registry update
yonga registry update [OPTIONS]
Fetch, verify and install a signed remote registry bundle. Only bundles signed by a key in ~/.config/yonga/registry-keys/ are accepted.
| Option |
Value |
Default |
Description |
--url |
text |
registry_url setting (unset) |
Bundle URL (https); the signature is fetched from URL.sig. |
--file |
path |
– |
A bundle already on disk. Cannot be combined with --url. |
--signature |
path |
FILE.sig |
Its signature. Only with --file. |
--yes, -y |
|
off |
Install without a prompt. |
--json |
|
off |
Emit the outcome as JSON. |
yonga registry update --file bundle.json
registry status
yonga registry status [OPTIONS]
Show the installed remote registry bundle and whether it still verifies. Exits 2 when an installed bundle is being ignored.
| Option |
Value |
Default |
Description |
--json |
|
off |
Emit the status as JSON. |
registry reset
yonga registry reset [OPTIONS]
Remove the installed remote bundle and go back to the built-in registry.
| Option |
Value |
Default |
Description |
--json |
|
off |
Emit the outcome as JSON. |
yonga env
yonga env [OPTIONS] COMMAND [ARGS]...
List and remove the managed environments exports run in. See Maintenance.
env list
Show the managed export environments and the disk they use.
| Option |
Value |
Default |
Description |
--json |
|
off |
Emit the environments as JSON. |
env remove
yonga env remove [OPTIONS] KEY
Delete one managed export environment. It is rebuilt if a later export needs it.
| Argument |
Description |
KEY |
Environment key, or a unique prefix. Required. |
| Option |
Value |
Default |
Description |
--yes, -y |
|
off |
Do not ask for confirmation. |
--json |
|
off |
Emit the result as JSON. |
yonga env remove fbfac3ac91ba
yonga cache
yonga cache [OPTIONS] COMMAND [ARGS]...
Show and prune the OpenVINO compile cache. yonga cache without a subcommand is the same as yonga cache list. See Maintenance.
| Option |
Value |
Default |
Description |
--json |
|
off |
Emit the entries as JSON. |
cache list
yonga cache list [OPTIONS]
List compile cache entries: size, last use, owner, and whether they are still used.
| Option |
Value |
Default |
Description |
--json |
|
off |
Emit the entries as JSON. |
cache prune
yonga cache prune [OPTIONS]
Remove orphaned and stale-runtime compile cache entries (and more, when asked).
| Option |
Value |
Default |
Description |
--older-than |
DAYS |
– |
Also remove any entry not used for this many days. |
--unattributed |
|
off |
Also remove entries that cannot be traced to any artifact. |
--all |
|
off |
Empty the whole compile cache. |
--dry-run |
|
off |
Show what would go; delete nothing. |
--yes, -y |
|
off |
Do not ask for confirmation. |
--json |
|
off |
Emit the outcome as JSON. |
yonga cache prune --dry-run
yonga self
yonga self [OPTIONS] COMMAND [ARGS]...
Put the yonga AppImage on PATH as 'yonga', or take it off again. Works only when yonga runs from the AppImage.
self install
yonga self install [OPTIONS]
Copy the running AppImage to ~/.local/bin/yonga so 'yonga' works in any shell.
| Option |
Value |
Default |
Description |
--bin-dir |
DIR |
~/.local/bin |
Directory to install the 'yonga' command in. |
--yes, -y |
|
off |
Replace an old AppImage without asking. |
--json |
|
off |
Emit the outcome as JSON. |
./yonga-0.6.0-x86_64.AppImage self install
self uninstall
yonga self uninstall [OPTIONS]
Remove the AppImage copy at ~/.local/bin/yonga. Models and caches are kept.
| Option |
Value |
Default |
Description |
--bin-dir |
DIR |
~/.local/bin |
Directory to remove the 'yonga' command from. |
--yes, -y |
|
off |
Do not ask for confirmation. |
--json |
|
off |
Emit the outcome as JSON. |
yonga connect
yonga connect [OPTIONS] COMMAND [ARGS]...
Register a prepared artifact with a client that talks to 'yonga serve'.
connect opencode
yonga connect opencode [OPTIONS] ARTIFACT
Add a prepared artifact to OpenCode's config as a 'yonga serve' provider, with the context and output limits read from the artifact. It shows the change and asks first, and refuses an artifact OpenCode cannot use: one without tool calling, or one whose prompt length is fixed under 8192 tokens, as every built-in NPU profile is. See Use it from OpenCode.
| Argument |
Description |
ARTIFACT |
Artifact id, repo id or path. Required. With --remove, the recorded id. |
| Option |
Value |
Default |
Description |
--remove |
|
off |
Take the connection back out; ARTIFACT is the recorded id. |
--host |
text |
127.0.0.1 |
Address 'yonga serve' listens on. |
--port |
integer, 1–65535 |
8000 |
Port 'yonga serve' listens on. |
--device |
text |
auto |
auto, npu, gpu or cpu, resolved as serve resolves it. |
--context |
integer, at least 1 |
– |
Context limit to write, instead of the one read from the artifact. |
--config |
PATH |
~/.config/opencode/opencode.json |
The OpenCode file to edit; $XDG_CONFIG_HOME/opencode/opencode.json when XDG_CONFIG_HOME is set. |
--yes, -y |
|
off |
Do not ask before writing. |
--json |
|
off |
Emit the outcome as JSON. |
yonga connect opencode 073d5d7c8ea75283 --port 8765