which local LLM does facet use？

[monithu]

rt

[ncoevoet]

Facet runs 100% locally — no cloud APIs. The "LLM" is the vision-language model used for tagging, captioning, and VLM critique, and which one loads depends on your VRAM profile:

Profile	Tagging / VLM	Hugging Face ID
legacy / 8gb	No LLM — CLIP similarity tagging	ViT-L-14 (open_clip)
16gb	Qwen3.5-2B	Qwen/Qwen3.5-2B
24gb	Qwen3.5-4B	Qwen/Qwen3.5-4B

A few other VLMs are also wired up and selectable in scoring_config.json if you want to swap: Qwen/Qwen2-VL-2B-Instruct, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-VL-2B-Instruct, and florence-community/Florence-2-large.

For context, the LLM is just one piece of a larger local model ensemble — none of these need an external service either:

• Embeddings / semantic search: SigLIP 2 NaFlex SO400M (google/siglip2-so400m-patch16-naflex) on 16gb/24gb, CLIP ViT-L-14 on legacy/8gb
• Aesthetic quality: TOPIQ (plus supplementary TOPIQ IAA, NR-Face, LIQE via PyIQA)
• Composition: SAMP-Net
• Faces: InsightFace buffalo_l
• Subject saliency: BiRefNet (ZhengPeng7/BiRefNet_dynamic)

So: no GPU at all → legacy/8gb profiles work with CLIP-based tagging and no generative LLM. With a GPU, you get the Qwen vision-language models for tags/captions/critique. See docs/CONFIGURATION.md for the full profile reference.