Nemotron 3 Embed 8B API

Nemotron 3 Embed 8B turns text into dense vectors for multilingual search and RAG, from $0.031.

4096 dimension vectors
33K token context
34 languages evaluated
OpenMDW 1.1, commercial use

from openai import OpenAI

client = OpenAI(
    base_url="https://api.modelsheep.com",
    api_key=API_KEY_MODELSHEEP,
)

response = client.embeddings.create(
    model="nvidia/Nemotron-3-Embed-8B-BF16",
    input="The quick brown fox jumps over the lazy dog.",
)

print(response.data[0].embedding)

How to use the Nemotron 3 Embed 8B API

1
Create a ModelSheep account and copy an API key from the dashboard.
2
Point the OpenAI SDK at the base url https://api.modelsheep.com and set the model to nvidia/Nemotron-3-Embed-8B-BF16.
3
Call /embeddings with your text in input, write query: in front of a search query and passage: in front of a document, and read a vector of 4096 numbers from the reply.

What Nemotron 3 Embed 8B is

Nemotron 3 Embed 8B is a text embedding model from NVIDIA. It uses a Ministral 3 8B encoder with bidirectional attention and average pooling, and it returns one dense vector per input string. NVIDIA built it for retrieval and semantic search, and evaluated it across 34 languages. The family also holds two smaller 1B checkpoints, and this is the accuracy-first one that NVIDIA reports at the top of the RTEB retrieval leaderboard on 16 July 2026.

What it does well

Builds the retrieval index for a multilingual question answering app.
Matches a query to a passage when the two are written in different languages.
Embeds a whole document in one call, up to the 33K limit.

What it is not for

Text only. A scanned page needs an OCR step before this model sees it.
No MTEB mean is published for it, so it cannot be lined up against embedding

Calling it

Call POST /embeddings with the model string nvidia/Nemotron-3-Embed-8B-BF16. The model expects a prefix on every string. Use query: for search queries and passage: for documents. The OpenAI-compatible endpoint does not add the prefix. Write the prefix into the strings you pass in input. To store a shorter vector, slice it from the front and re-normalize it with L2 yourself.

nvidia/Nemotron-3-Embed-8B-BF16 /embeddings

SDKs and languages

Nemotron 3 Embed 8B answers on an HTTP endpoint, so any language that can post JSON reaches it. These are the ones the box above writes out.

OpenAI SDK: PythonTypeScriptGoJavaC#PHPRust
HTTP: cURLPythonTypeScriptGoRustJavaC#PHP

Specifications

Author	NVIDIA
Category	embedding
Accepts	text
Returns	vector
Total parameters	7.953B
Active parameters	dense, all of them
Score	not rated
Languages	not publishedNVIDIA publishes no support claim. The card reports evaluation across 34 languages, which is a different thing.
Precision served	FP8
Context	33K tokens
Vector	4096 dimensions
License	OpenMDW 1.1

Price

Input

$0.031

per million tokens

Output

n/a

this model returns a vector

Against OpenRouter

11%

below the cheapest of 1 provider

The cheapest OpenRouter price for this model was $0.035 in and n/a out, with a median of $0.035 and n/a across the comparable providers. Read on 2026-09-17.

License

NVIDIA releases the weights under OpenMDW 1.1, a permissive license that allows commercial use and imposes no restriction on the vectors the model returns. Redistributing the weights requires keeping a copy of the license and origin notices. The grant ends if you sue over the model materials. Read the text before you ship a copy.

Parameters: 7.6B
Context: 33K
Vector: 4096
Input per 1M: $0.009
Output per 1M: n/a
License: Apache 2.0