Command-line interface for running AI models locally on Qualcomm chipsets. Interfaces with the GenieX core runtime and supports two inference backends: QAIRT and llama.cpp.

GENIEX_LOG controls log output across the CLI, the C/C++ SDK, and all language bindings (Go, Python, Android):

export GENIEX_LOG="debug"          # bash / zsh
$env:GENIEX_LOG="debug"            # PowerShell

The --log flag is equivalent and takes precedence over GENIEX_LOG when both are set:

geniex --log debug list

NO_COLOR=1 disables ANSI colors.

This also governs the qairt backend's own QNN logging: at none (the default), QNN emits nothing extra; any other level asks QNN for full verbosity and lets the same threshold decide what's shown.

The qairt backend has a fixed context length (e.g. 4096 tokens). By default, once the accumulated conversation history plus a new prompt exceeds it, geniex infer returns an out-of-context error and the session cannot continue.

Pass --sliding-window to opt into evicting the oldest tokens (above a small anchored prefix) instead, letting the conversation continue past the context limit:

geniex infer <model> --sliding-window

This sets sliding_window: true on the generation config for every generate() call; llama_cpp ignores it (it always context-shifts). Without the flag, exceeding the context length still returns the original error — the feature is strictly opt-in.

Pull a model non-interactively:

geniex pull <model>[:<precision>] --model-type <model-type>

Pull from a specific model hub:

geniex pull <model>
geniex pull <model> --model-hub aihub   # options: aihub, hf, localfs

Import a model from the local filesystem:

# hf download <model> --local-dir /path/to/modeldir
geniex pull <model> --model-hub localfs --local-path /path/to/modeldir