oir-vendor-models — permissively-licensed model bundle

Reference baseline models that ship with a JibarOS Cuttlefish build. Model binaries are downloaded by tools/fetch-models.sh from canonical upstream homes (HuggingFace), not vendored as Git LFS — keeps the repo lightweight. Installs to /product/etc/oir/ via prebuilt_etc in Android.bp.

What's included

File	Capability	License	Size
qwen2.5-0.5b-instruct-q4_k_m.gguf	text.complete / text.translate	Apache 2.0	~470 MB
all-MiniLM-L6-v2.Q8_0.gguf	text.embed	Apache 2.0	~24 MB
whisper-tiny-en.Q5.bin	audio.transcribe	MIT	~31 MB
voice-sample.wav	OirDemo audio.transcribe demo input	CC0 (sonic/talking.wav resampled)	~720 KB

See NOTICE for full attribution.

Deliberately not included

These capabilities are declared in capabilities.xml with no platform default — OEMs bake their choice:

Capability	Why
vision.describe	VLMs are typically >500 MB; OEMs pick size/perf tradeoff (SmolVLM-500M to LLaVA-7B)
vision.detect	YOLO family is AGPL; OEMs accept obligations or pick RT-DETR (Apache)
audio.synthesize	Piper voices need locale-specific G2P sidecars; no universal default
vision.embed	siglip-base-patch16-224 is declared in Android.bp for OEMs who want to opt in (PRODUCT_PACKAGES += oir_siglip_model); not installed by default — ~372 MB is more than the platform-default story warrants

Git LFS

Model weights are tracked via Git LFS. To clone:

git lfs install
git clone https://github.com/jibar-os/oir-vendor-models

Wiring up

In a JibarOS tree, add to your device PRODUCT_PACKAGES:

PRODUCT_PACKAGES += \
    oir_default_model \
    oir_minilm_model \
    oir_whisper_tiny_en_model \
    oir_siglip_model \
    oir_voice_sample_wav

The reference Cuttlefish device tree (device_google_cuttlefish) already does this.

See also

github.com/Jibar-OS/JibarOS — why we bundle what we bundle + OEM bake-in guide.