Skip to main content

Nemotron 3.5 Embed 1B

Registry ids: nemotron-3.5-embed-1b (hosted NIM), nemotron-3.5-embed-1b-local (self-hosted OpenAI /v1/embeddings).

Model id: nvidia/nemotron-3.5-embed-1b (build.nvidia.com).

Hosted (NIM)​

export NVIDIA_API_KEY=...
opentryon embed --model nemotron-3.5-embed-1b \
--text "linen trench coat" "wool blazer" \
--query "lightweight outerwear"
opentryon embed --model nemotron-3.5-embed-1b \
--image page1.png catalog.png \
--input-type passage --query "summer dress SKU layout"

Local (NeMo Retriever Embedding NIM / vLLM)​

Serve the BF16 or NVFP4 checkpoint with NVIDIA’s Embedding NIM, then:

export NEMOTRON_EMBED_BASE_URL=http://127.0.0.1:8000/v1
opentryon embed --model nemotron-3.5-embed-1b-local --text "hello" --query "hi"

CLI flags​

FlagPurpose
--input-typequery vs passage (retrieval)
--modalitytext, image, or text_image when mixing inputs
--dimOptional dynamic embedding size
--longest-edgeImage resize cap (256–1540)

Vectors are L2-normalized and saved to .npz like other embed models.