Nemotron-3.5-Embed-8B-BF16
by nvidia
NVIDIA's 8.38B multimodal embedder: search document pages as images, text or both from a text query, with 4,096-dimensional vectors in the same text space as Nemotron-3-Embed-8B
nvidia/Nemotron-3.5-Embed-8B-BF16Deploy Nemotron-3.5-Embed-8B-BF16
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
Nemotron-3.5-Embed-8B-BF16 is an embedding model from NVIDIA, published on Hugging Face in October 2026 under the OpenMDW-1.1 license, which allows commercial use. It embeds a document page as an image, as text, or as both together, and embeds queries as text, so a question can find the page that answers it when the answer sits in a table, a chart or an infographic. Each input becomes one 4,096-dimensional vector.
It is the larger sibling of Nemotron-3.5-Embed-1B and the model the 1B was distilled from. Its language model was initialized from Nemotron-3-Embed-8B and kept frozen while only the vision encoder and projector were trained, so the card says the two models produce identical text embeddings. A corpus already indexed with Nemotron-3-Embed-8B can add page images without re-embedding its text.
NVIDIA evaluated text retrieval in 34 languages and image retrieval in 18, on ViDoRe V3 and MIRACL-VISION with NDCG@10. The card names those benchmarks without publishing scores, and we have not measured it.
Architecture
A vision encoder initialized from Mistral's Ministral-3-8B-Instruct-2512 and a language model initialized from nvidia/Nemotron-3-Embed-8B-BF16, 8.38B parameters in total, BF16 weights. Token-level hidden states are mean-pooled and L2-normalized into one 4,096-dimensional vector per input; the vector can be truncated to 1,024 or 512 dimensions and re-normalized. Context is 32,768 tokens, including image tokens; page images default to a 1,120-pixel longest edge and can go up to 1,540. Training kept the language model frozen and fine-tuned the vision encoder and multimodal projector on 2.3M image samples, with 50M+ text-only samples listed in the training data. Inputs take a query: or passage: prefix.
Mixpeek SDK Integration
# Embed each PDF page image (with its OCR text when you have it) with
# Nemotron-3.5-Embed-8B-BF16, then upsert the vectors into a Mixpeek Vector Store
# namespace you query directly. Needs a CUDA GPU (A100 80GB, H100 80GB or GB200 per the card).
import torch
from PIL import Image
from sentence_transformers import SentenceTransformer
from mixpeek import Mixpeek
model = SentenceTransformer(
"nvidia/Nemotron-3.5-Embed-8B-BF16",
device="cuda:0",
model_kwargs={"attn_implementation": "flash_attention_2", "dtype": torch.bfloat16},
)
client = Mixpeek(api_key="YOUR_KEY")
pages = [{"id": "q3-report-p14", "image": "q3-report-p14.png", "text": "Revenue by region..."}]
vectors = model.encode_document(
[{"image": Image.open(p["image"]), "text": p["text"]} for p in pages]
)
client.namespaces.upsert(
namespace="reports",
documents=[{"id": p["id"], "dense_embedding": v.tolist(), "text": p["text"]}
for p, v in zip(pages, vectors)],
)Capabilities
- One 4,096-dimensional vector per page image, text chunk, or image plus text
- Text queries against pages that hold tables, charts and infographics
- Truncation to 1,024 or 512 dimensions for cheaper storage
- Identical text embeddings to Nemotron-3-Embed-8B-BF16
- Sentence Transformers, Transformers and vLLM; OpenMDW-1.1 license
Use Cases on Mixpeek
Performance
The card lists A100 80GB, H100 80GB and GB200 186GB as supported GPUs and uses Flash Attention 2 in its examples. A 4,096-dimensional float32 vector is 16 KB, four times the 1B sibling's 2,048 dimensions at 8 KB, so truncation matters at scale. No scores are published on the card and we have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is Nemotron-3.5-Embed-8B-BF16?
An NVIDIA embedding model with 8.38B parameters that turns a document page (as an image, text, or both) or a text query into one 4,096-dimensional vector, for retrieval over PDFs, slides and scanned documents. It is released under the OpenMDW-1.1 license, which allows commercial use.
Should I use the 8B or the 1B Nemotron-3.5-Embed?
The 1B was distilled from the 8B, so the 8B is the teacher and the 1B is the cheaper model to run: 1.55B parameters and 2,048 dimensions against 8.38B and 4,096. NVIDIA publishes no scores for either on the cards, so compare them on a sample of your own pages and queries before choosing.
Do I need to re-embed my Nemotron-3-Embed-8B text index?
According to the card, no. The language model was initialized from Nemotron-3-Embed-8B-BF16 and frozen during training, and the card says the two produce identical text embeddings. You can add page-image vectors from the new model next to existing text vectors.
Can queries be images?
No. The card says queries are text only. Documents can be an image, text, or an image and text together.
How do I use Nemotron-3.5-Embed-8B with Mixpeek?
Embed your page images and queries with it on your own GPU and upsert the vectors into a Mixpeek Vector Store namespace, as in the example on this page. MVS can combine those dense vectors with BM25 over the page text in one hybrid query. On Enterprise, the weights can also be uploaded and run by a custom plugin.
Specification
Build a pipeline with Nemotron-3.5-Embed-8B-BF16
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free