Nemotron-3.5-Embed 1B & 8B

Ollama builds of NVIDIA’s Nemotron-3.5-Embed family: multilingual multimodal (text + image) embedding models. Converted to GGUF with llama.cpp, then quantized. Both sizes expose the same Ollama API and accept text, images, or text + image.

• Nemotron-3.5-Embed-1B-BF16
• Nemotron-3.5-Embed-8B-BF16

Multimodal. Every tag bundles its BF16 Pixtral vision projector (mmproj), so ollama show reports the vision capability and /api/embed accepts image input via map-shaped entries: {"text": "...", "image": "<base64>"}. Text-only inputs work as before.

Tags

Sizes are what ollama list reports and include the bundled projector.

1B

All include the BF16 Pixtral projector (clip, 413.79M parameters).

8B

All include the BF16 Pixtral projector (clip, 428.48M parameters).

The :8b alias points at Q4_K_M, the same way :1b points at Q8_0.

Model details

1B

• Parameters: 1.55B total (the text tower is ~1.1B in the GGUF)
• Text architecture: Ministral-3 (16 layers, hidden 2048, 24 attention heads / 8 KV heads)
• Vision projector: Pixtral, 413.79M parameters, 24 blocks, image size 1540 px (preproc 1120 px), projection dim 2048
• Embedding: 2048-dim, mean-pooled, L2-normalized
• Context: 32,768 tokens
• Languages: 34 (en, ar, as, bn, bg, zh, da, nl, fi, fr, de, hi, id, it, ja, ko, ms, mr, ne, no, fa, pt, ro, ru, es, sw, sv, ta, te, th, uk, ur, vi)

8B

• Parameters: 8.38B total
• Text architecture: Ministral-3 (34 layers, hidden 4096, 32 attention heads / 8 KV heads, head_dim 128, FFN 14336, vocab 131072)
• Vision projector: Pixtral, 428.48M parameters, 24 blocks, image size 1540 px (preproc 1120 px), projection dim 4096
• Embedding: 4096-dim, mean-pooled, L2-normalized
• Context: 32,768 tokens
• Languages: 34 for text (same list as 1B), 18 for images
• The text tower is the frozen Nemotron-3-Embed-8B model, so it shares the embedding space of that text-only model

Both sizes support dynamic embedding sizes: truncate the vector from the end (for example to 1024 or 512 dimensions) and re-normalize. Truncated embeddings remain effective for retrieval.

Choosing a size

Do not mix sizes in one index. A 1B vector and an 8B vector are not comparable. Index and query with the same tag.

Usage

CLI

ollama run maternion/nemotron-3.5-embed:1b "query: What is NVIDIA NeMo Framework?"
ollama run maternion/nemotron-3.5-embed:8b "query: What is NVIDIA NeMo Framework?"

This prints the raw vector. Use the HTTP API for real workloads.

Use the model’s task prefixes:

• queries: query:
• documents: passage:

HTTP API

Docs: https://docs.ollama.com/api/embed

Single input:

curl http://localhost:11434/api/embed -d '{
  "model": "maternion/nemotron-3.5-embed:8b",
  "input": "query: What is NVIDIA NeMo Framework?"
}'

Batch of queries and documents:

curl http://localhost:11434/api/embed -d '{
  "model": "maternion/nemotron-3.5-embed:8b",
  "input": [
    "query: What is NVIDIA NeMo Framework?",
    "passage: NeMo Framework offers comprehensive functionalities for developing both LLMs and Multimodal Models.",
    "passage: A biological foundation model designed to analyze and generate DNA, RNA, and protein sequences."
  ]
}'

Python with cosine similarity (swap :1b for :8b as needed):

import requests
import numpy as np

URL = "http://localhost:11434/api/embed"
MODEL = "maternion/nemotron-3.5-embed:8b"   # or :1b


def embed(inputs):
    r = requests.post(URL, json={"model": MODEL, "input": inputs})
    r.raise_for_status()
    return np.array(r.json()["embeddings"], dtype=np.float32)


vectors = embed([
    "query: What is NVIDIA NeMo Framework?",
    "passage: NeMo Framework offers comprehensive functionalities for developing both LLMs and Multimodal Models.",
    "passage: A biological foundation model designed to analyze and generate DNA, RNA, and protein sequences.",
])
print(vectors.shape)
scores = vectors[0] @ vectors[1:].T
print(scores)

Truncated embeddings:

curl http://localhost:11434/api/embed -d '{
  "model": "maternion/nemotron-3.5-embed:8b",
  "input": "query: What is NVIDIA NeMo Framework?",
  "dimensions": 1024
}'

Image embeddings (Ollama API)

/api/embed accepts map-shaped entries with any of text, image, or audio (base64). Image + query:

IMG=$(base64 -w0 screenshot.png)
curl http://localhost:11434/api/embed -d "{
  \"model\": \"maternion/nemotron-3.5-embed:8b\",
  \"input\": {\"text\": \"query: What is shown in this screenshot?\", \"image\": \"$IMG\"}
}"

Or from Python, including image-only and a mixed batch:

import base64
import requests

URL = "http://localhost:11434/api/embed"
MODEL = "maternion/nemotron-3.5-embed:8b"   # or :1b
b64 = base64.b64encode(open("screenshot.png", "rb").read()).decode()


def embed(inp):
    r = requests.post(URL, json={"model": MODEL, "input": inp})
    r.raise_for_status()
    return r.json()["embeddings"]


print(len(embed({"text": "query: What is shown?", "image": b64})[0]))
embed({"image": b64})                                                  # image only

vecs = embed([                                                         # mixed batch
    "query: What is shown?",
    {"image": b64},
    {"text": "passage: a screenshot", "image": b64},
])
print(len(vecs))  # 3

Docs: https://docs.ollama.com/api/embed. The Go API (Input) documents the accepted shapes: a string, []string, or maps with text/image/audio.

Image embeddings (llama.cpp server, alternative)

The same GGUFs also run under llama.cpp if you are not using Ollama. Files are on Hugging Face:

• Nemotron-3.5-Embed-1B-GGUF
• Nemotron-3.5-Embed-8B-GGUF

1B:

./build/bin/llama-server \
  -m Nemotron-3.5-Embed-1B-Q8_0.gguf \
  -mm mmproj-Nemotron-3.5-Embed-1B-BF16.gguf \
  --embedding --pooling mean --ctx-size 8192 -ngl 0 --port 8099

8B:

./build/bin/llama-server \
  -m Nemotron-3.5-Embed-8B-Q8_0.gguf \
  -mm mmproj-Nemotron-3.5-Embed-8B-BF16.gguf \
  --embedding --pooling mean --ctx-size 8192 -ngl 0 --port 8099

Then send an OpenAI-style content array — image first, then the query text:

curl http://127.0.0.1:8099/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "input": [{
      "content": [
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64>"}},
        {"type": "text", "text": "query: What is shown in this screenshot?"}
      ]
    }]
  }'

Or from Python:

import base64
import requests

b64 = base64.b64encode(open("screenshot.png", "rb").read()).decode()
r = requests.post("http://127.0.0.1:8099/v1/embeddings", json={
    "input": [{"content": [
        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}},
        {"type": "text", "text": "query: What is shown in this screenshot?"},
    ]}]
})
print(len(r.json()["data"][0]["embedding"]))

The legacy non-OAI endpoint is POST /embedding, which takes the { "prompt_string": "... <__media__> ...", "multimodal_data": ["<base64>"] } form (see the llama.cpp server docs).

For the official vLLM /v2/embed image API, see the upstream model cards: 1B and 8B.

Benchmarks

NDCG@10:

Published BF16 scores; quantized variants may differ slightly. These benchmarks include image retrieval, which the bundled projector supports through the map-shaped /api/embed input.

License

OpenMDW-1.1. Base models: Apache 2.0.

Links

• 1B model: https://huggingface.co/nvidia/Nemotron-3.5-Embed-1B-BF16
• 8B model: https://huggingface.co/nvidia/Nemotron-3.5-Embed-8B-BF16
• 1B GGUF: https://huggingface.co/maternion/Nemotron-3.5-Embed-1B-GGUF
• 8B GGUF: https://huggingface.co/maternion/Nemotron-3.5-Embed-8B-GGUF
• Ollama embeddings API: https://docs.ollama.com/api/embed