Model Architecture:
Architecture Type: Transformer
Network Architecture: Ministral-3-3B-Instruct-2512 based pruned model
Number of model parameters: The model has approximately 1.14B parameters
Hidden Size: 2048
The Nemotron-3-Embed-1B-BF16 model is a transformer-based text embedding model trained with bidirectional attention masking, where the final embedding vector is obtained by applying average pooling to the transformer’s token-level representations. It encodes each input text into a dense embedding vector of dimension 2048.
The Nemotron-3-Embed-1B-BF16 was derived from the nemotron-3-embed-3b text-embedding model through two iterative rounds of structured pruning and distillation. First, the 3B parent model was pruned to 2B using NVIDIA ModelOpt mcore_minitron Neural Architecture Search (NAS) [NVIDIA/Model-Optimizer, paper]. This process searches across hidden width, FFN size, attention heads, and depth, then selects the best candidate from the top-10 Pareto front. Candidates were evaluated against the parent model’s representations using a 50k in-domain calibration corpus, which was also used to estimate importance scores.
The resulting 2B model was then distilled from the fine-tuned Nemotron-3-Embed-8B-BF16 embedding teacher model to recover accuracy. Distillation used combined cosine distance loss (COS) and mean squared error (MSE) loss on a multilingual, in-domain retrieval data blend. The same pruning-and-distillation procedure, using the same dataset blend, was then repeated to produce the final 1.14B embedding model.
Input Type(s): Text
Input Format(s):
Input Parameters:
- Text: One-Dimensional (1D)
Other Properties Related to Input: Text inputs should be tokenized by the model tokenizer. The model’s max sequence length is 32768. Longer inputs should be chunked or truncated.
Output(s)
Output Type(s): Floats
Output Format(s):
Output Parameters: One-Dimensional (1D) embedding vector per input text string
Other Properties Related to Output: The model outputs a 2048-dimensional embedding vector for each input text string.
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
Usage
For local Python examples, use the Hugging Face model ID by default. If you are working from a local checkout, replace MODEL_ID with that path. For vLLM online serving from a local checkpoint, use the local checkpoint example.
MODEL_ID = "nvidia/Nemotron-3-Embed-1B-BF16"
Use this model for retrieval-style embeddings. Add the query: prefix for queries and the passage: prefix for documents. Embeddings are L2-normalized, so dot product and cosine similarity are equivalent. The output tables use q[i] for queries and d[i] for documents. Scores are rounded to four decimals. Exact values can vary by runtime and package version.
Local Python Dependencies
The BF16 checkpoints support Transformers 5.2.0 and above. The examples also require a CUDA-enabled PyTorch installation that matches your driver and CUDA environment.
If you are not using an NVIDIA PyTorch container, install PyTorch first. Use the PyTorch local installation selector to choose the command that matches your operating system, package manager, and CUDA environment. If the default PyPI wheel matches your CUDA environment, run:
pip install --upgrade torch
For non-default CUDA environments, the selector can include an additional --index-url argument. Use that CUDA-specific index URL when the default PyPI wheel does not match your CUDA environment.
Then install Transformers and Sentence Transformers:
pip install --upgrade "transformers>=5.2.0" "sentence-transformers>=5.4.1"
NVIDIA tested the examples in nvcr.io/nvidia/pytorch:26.06-py3 with the container-provided Torch and CUDA stack. Inside NVIDIA PyTorch containers, do not upgrade Torch. Install only the missing packages:
pip install --upgrade "transformers>=5.2.0" "sentence-transformers>=5.4.1"
The tested nvcr.io/nvidia/pytorch:26.06-py3 container includes flash-attn, so the snippets use FlashAttention-2 by default. If your environment does not have FlashAttention-2, set attn_implementation or ATTN_IMPLEMENTATION to sdpa.
Use Sentence Transformers for the simplest local Python interface. It reads the saved query and document prompts and normalization metadata.
import torch
from sentence_transformers import SentenceTransformer
MODEL_ID = "nvidia/Nemotron-3-Embed-1B-BF16"
model = SentenceTransformer(
MODEL_ID,
device="cuda",
model_kwargs={
"dtype": torch.bfloat16,
"attn_implementation": "flash_attention_2",
},
processor_kwargs={"padding_side": "left"},
)
model.max_seq_length = 32768
QUERIES = [
"Write a Python function that counts the frequency of each element in a list of lists.",
"Write a function that orders a dictionary with tuple keys by the product of each key's tuple values.",
"What symptoms and common triggers help distinguish eczema from other inflammatory skin conditions?",
"How can someone reduce exposure to pollen during allergy season?",
]
DOCUMENTS = [
"def frequency_lists(list1):\n flattened = [item for sublist in list1 for item in sublist]\n counts = {}\n for item in flattened:\n if item in counts:\n counts[item] += 1\n else:\n counts[item] = 1\n return counts",
"def sort_dict_item(test_dict):\n return {key: test_dict[key] for key in sorted(test_dict.keys(), key=lambda ele: ele[0] * ele[1])}",
"Eczema commonly causes itchy, dry, inflamed patches of skin. The affected areas may look red, scaly, cracked, or darker than the surrounding skin depending on skin tone. Symptoms can flare after exposure to irritants, allergens, stress, or changes in weather.",
"People with pollen allergy can reduce exposure by staying indoors on dry, windy days, avoiding early-morning outdoor activity, and going outside after rain when pollen levels are lower. They should check pollen forecasts, close windows and doors when counts are high, and consider starting allergy medication before symptoms begin if high pollen is expected. After being outside, showering, changing clothes, avoiding outdoor laundry drying, and wearing a face mask for yard work can help limit pollen contact.",
]
query_embeddings = model.encode_query(QUERIES, batch_size=8, convert_to_tensor=True)
document_embeddings = model.encode_document(DOCUMENTS, batch_size=8, convert_to_tensor=True)
scores = model.similarity(query_embeddings, document_embeddings)
print("Similarity scores:")
print(f"{'':>8}" + "".join(f"d[{i}] " for i in range(scores.shape[1])))
for query_index, row in enumerate(scores):
print(f"q[{query_index}] " + " ".join(f"{score.item():>7.4f}" for score in row))
Similarity scores:
d[0] d[1] d[2] d[3]
q[0] 0.8125 0.0255 0.0005 -0.0312
q[1] 0.0447 0.6484 -0.0520 0.0386
q[2] -0.0095 -0.0410 0.6484 0.1006
q[3] -0.0216 0.0214 0.1211 0.7734
Use Transformers when you want manual control over tokenization, pooling, or batching.
import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer
MODEL_ID = "nvidia/Nemotron-3-Embed-1B-BF16"
MAX_LENGTH = 32768
BATCH_SIZE = 8
DTYPE = torch.bfloat16
ATTN_IMPLEMENTATION = "flash_attention_2"
QUERIES = [
"Write a Python function that counts the frequency of each element in a list of lists.",
"Write a function that orders a dictionary with tuple keys by the product of each key's tuple values.",
"What symptoms and common triggers help distinguish eczema from other inflammatory skin conditions?",
"How can someone reduce exposure to pollen during allergy season?",
]
DOCUMENTS = [
"def frequency_lists(list1):\n flattened = [item for sublist in list1 for item in sublist]\n counts = {}\n for item in flattened:\n if item in counts:\n counts[item] += 1\n else:\n counts[item] = 1\n return counts",
"def sort_dict_item(test_dict):\n return {key: test_dict[key] for key in sorted(test_dict.keys(), key=lambda ele: ele[0] * ele[1])}",
"Eczema commonly causes itchy, dry, inflamed patches of skin. The affected areas may look red, scaly, cracked, or darker than the surrounding skin depending on skin tone. Symptoms can flare after exposure to irritants, allergens, stress, or changes in weather.",
"People with pollen allergy can reduce exposure by staying indoors on dry, windy days, avoiding early-morning outdoor activity, and going outside after rain when pollen levels are lower. They should check pollen forecasts, close windows and doors when counts are high, and consider starting allergy medication before symptoms begin if high pollen is expected. After being outside, showering, changing clothes, avoiding outdoor laundry drying, and wearing a face mask for yard work can help limit pollen contact.",
]
def average_pool(last_hidden_state: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
last_hidden = last_hidden_state.masked_fill(~attention_mask[..., None].bool(), 0.0)
return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
if not torch.cuda.is_available():
raise RuntimeError("CUDA is required for practical BF16 inference.")
device = torch.device("cuda")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, padding_side="left")
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model = AutoModel.from_pretrained(
MODEL_ID,
dtype=DTYPE,
attn_implementation=ATTN_IMPLEMENTATION,
).to(device)
model.eval()
def encode_texts(texts: list[str]) -> torch.Tensor:
embedding_batches = []
for start in range(0, len(texts), BATCH_SIZE):
encoded = tokenizer(
texts[start : start + BATCH_SIZE],
max_length=MAX_LENGTH,
truncation=True,
padding=True,
return_tensors="pt",
)
encoded = {name: tensor.to(device) for name, tensor in encoded.items()}
with torch.inference_mode():
output = model(**encoded)
pooled = average_pool(output.last_hidden_state, encoded["attention_mask"])
embeddings = F.normalize(pooled, p=2, dim=-1)
embedding_batches.append(embeddings.detach().cpu().to(torch.float32))
return torch.cat(embedding_batches, dim=0)
embeddings = encode_texts(
["query: " + query for query in QUERIES]
+ ["passage: " + doc for doc in DOCUMENTS]
)
query_embeddings = embeddings[: len(QUERIES)]
document_embeddings = embeddings[len(QUERIES) :]
scores = query_embeddings @ document_embeddings.T
print("Similarity scores:")
print(f"{'':>8}" + "".join(f"d[{i}] " for i in range(scores.shape[1])))
for query_index, row in enumerate(scores):
print(f"q[{query_index}] " + " ".join(f"{score.item():>7.4f}" for score in row))
Similarity scores:
d[0] d[1] d[2] d[3]
q[0] 0.8069 0.0252 0.0001 -0.0312
q[1] 0.0446 0.6466 -0.0514 0.0385
q[2] -0.0098 -0.0410 0.6450 0.0998
q[3] -0.0215 0.0212 0.1197 0.7679
vLLM Dependencies
For BF16, use vllm==0.25.0 for /v2/embed serving. NVIDIA also validated vllm serve "$MODEL_ID" with vllm/vllm-openai:v0.20.0 through v0.24.0 for this checkpoint.
pip install --upgrade "vllm==0.25.0" openai requests numpy
vLLM Offline Python
Use the offline Python API when you want local vLLM inference without running an HTTP server. LLM.embed accepts formatted strings, so add the query: and passage: prefixes manually.
import numpy as np
from vllm import LLM
MODEL_ID = "nvidia/Nemotron-3-Embed-1B-BF16"
QUERIES = [
"Write a Python function that counts the frequency of each element in a list of lists.",
"Write a function that orders a dictionary with tuple keys by the product of each key's tuple values.",
"What symptoms and common triggers help distinguish eczema from other inflammatory skin conditions?",
"How can someone reduce exposure to pollen during allergy season?",
]
DOCUMENTS = [
"def frequency_lists(list1):\n flattened = [item for sublist in list1 for item in sublist]\n counts = {}\n for item in flattened:\n if item in counts:\n counts[item] += 1\n else:\n counts[item] = 1\n return counts",
"def sort_dict_item(test_dict):\n return {key: test_dict[key] for key in sorted(test_dict.keys(), key=lambda ele: ele[0] * ele[1])}",
"Eczema commonly causes itchy, dry, inflamed patches of skin. The affected areas may look red, scaly, cracked, or darker than the surrounding skin depending on skin tone. Symptoms can flare after exposure to irritants, allergens, stress, or changes in weather.",
"People with pollen allergy can reduce exposure by staying indoors on dry, windy days, avoiding early-morning outdoor activity, and going outside after rain when pollen levels are lower. They should check pollen forecasts, close windows and doors when counts are high, and consider starting allergy medication before symptoms begin if high pollen is expected. After being outside, showering, changing clothes, avoiding outdoor laundry drying, and wearing a face mask for yard work can help limit pollen contact.",
]
def main():
llm = LLM(model=MODEL_ID)
texts = ["query: " + query for query in QUERIES] + [
"passage: " + doc for doc in DOCUMENTS
]
outputs = llm.embed(texts, use_tqdm=False)
embeddings = np.array(
[output.outputs.embedding for output in outputs],
dtype=np.float32,
)
query_embeddings = embeddings[: len(QUERIES)]
document_embeddings = embeddings[len(QUERIES) :]
scores = query_embeddings @ document_embeddings.T
print("Similarity scores:")
print(f"{'':>8}" + "".join(f"d[{i}] " for i in range(scores.shape[1])))
for query_index, row in enumerate(scores):
print(f"q[{query_index}] " + " ".join(f"{score:>7.4f}" for score in row))
if __name__ == "__main__":
main()
Similarity scores:
d[0] d[1] d[2] d[3]
q[0] 0.8110 0.0255 0.0003 -0.0312
q[1] 0.0448 0.6469 -0.0517 0.0386
q[2] -0.0100 -0.0400 0.6469 0.1003
q[3] -0.0224 0.0217 0.1199 0.7692
vLLM Online Serving
MODEL_ID=nvidia/Nemotron-3-Embed-1B-BF16
vllm serve "$MODEL_ID"
vLLM defaults to port 8000. Add host and port when you need an explicit bind address or a non-default port:
vllm serve "$MODEL_ID" --host 0.0.0.0 --port 8000
If you serve the Hugging Face model ID, the served model name defaults to nvidia/Nemotron-3-Embed-1B-BF16. If you serve a local checkpoint path, vLLM still reads the model config and weights from that path; --served-model-name only sets the model name accepted by API requests.
MODEL_PATH=/path/to/local/Nemotron-3-Embed-1B-BF16
vllm serve "$MODEL_PATH" --host 0.0.0.0 --port 8000 --served-model-name nvidia/Nemotron-3-Embed-1B-BF16
With --served-model-name, client requests continue to use MODEL = "nvidia/Nemotron-3-Embed-1B-BF16". If you omit it, use the served name reported by /v1/models in client requests.
Recommended Retrieval Endpoint
After the server is running, use /v2/embed for retrieval. Send raw query and document strings. input_type applies the saved query and document prompts:
import numpy as np
import requests
MODEL = "nvidia/Nemotron-3-Embed-1B-BF16"
URL = "http://localhost:8000/v2/embed"
QUERIES = [
"Write a Python function that counts the frequency of each element in a list of lists.",
"Write a function that orders a dictionary with tuple keys by the product of each key's tuple values.",
"What symptoms and common triggers help distinguish eczema from other inflammatory skin conditions?",
"How can someone reduce exposure to pollen during allergy season?",
]
DOCUMENTS = [
"def frequency_lists(list1):\n flattened = [item for sublist in list1 for item in sublist]\n counts = {}\n for item in flattened:\n if item in counts:\n counts[item] += 1\n else:\n counts[item] = 1\n return counts",
"def sort_dict_item(test_dict):\n return {key: test_dict[key] for key in sorted(test_dict.keys(), key=lambda ele: ele[0] * ele[1])}",
"Eczema commonly causes itchy, dry, inflamed patches of skin. The affected areas may look red, scaly, cracked, or darker than the surrounding skin depending on skin tone. Symptoms can flare after exposure to irritants, allergens, stress, or changes in weather.",
"People with pollen allergy can reduce exposure by staying indoors on dry, windy days, avoiding early-morning outdoor activity, and going outside after rain when pollen levels are lower. They should check pollen forecasts, close windows and doors when counts are high, and consider starting allergy medication before symptoms begin if high pollen is expected. After being outside, showering, changing clothes, avoiding outdoor laundry drying, and wearing a face mask for yard work can help limit pollen contact.",
]
def embed(input_type: str, texts: list[str]) -> np.ndarray:
response = requests.post(
URL,
json={
"model": MODEL,
"input_type": input_type,
"texts": texts,
"embedding_types": ["float"],
"truncate": "END",
},
timeout=120,
)
response.raise_for_status()
return np.array(response.json()["embeddings"]["float"], dtype=np.float32)
query_embeddings = embed("query", QUERIES)
document_embeddings = embed("document", DOCUMENTS)
scores = query_embeddings @ document_embeddings.T
print("Similarity scores:")
print(f"{'':>8}" + "".join(f"d[{i}] " for i in range(scores.shape[1])))
for query_index, row in enumerate(scores):
print(f"q[{query_index}] " + " ".join(f"{score:>7.4f}" for score in row))
Similarity scores:
d[0] d[1] d[2] d[3]
q[0] 0.8109 0.0252 0.0001 -0.0314
q[1] 0.0448 0.6470 -0.0516 0.0387
q[2] -0.0103 -0.0400 0.6470 0.1000
q[3] -0.0223 0.0217 0.1199 0.7693
You can also use the OpenAI-compatible /v1/embeddings endpoint. For those requests, pass strings in input and manually prefix them with query: or passage: .
Software Integration:
Runtime Engine(s): PyTorch, vLLM
Supported Hardware Microarchitecture Compatibility:
- NVIDIA Ampere
- NVIDIA Blackwell
- NVIDIA Hopper
Supported Operating System(s):
The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
Model Version(s):
Nemotron-3-Embed-1B-BF16
Training, Testing, and Evaluation Datasets:
Dataset Overview:
Total Size: 8.5M+ data points
Total Number of Datasets: 161 dataset files
Dataset Partition: Training [100%], Testing [N/A — evaluation benchmarks used separately], Validation [N/A — evaluation benchmarks used separately].
Model distillation training was conducted using publicly available, commercially permissible datasets and synthetically generated datasets. Synthetic data was created either by generating queries from seed documents or by generating complete question–answer pairs through LLM-based prompting using the models listed below.
Public Datasets:
Synthetic Datasets:
Synthetic query-document pairs were generated either from scratch or by using seed datasets to generate queries with the models listed below.
Training Dataset:
Data Modality:
Training Data Size:
Text Training Data Size: 8.5M+
Data Collection Method by dataset: Hybrid: Human, Automated, Synthetic
Labeling Method by dataset: Hybrid: Human, Automated, Synthetic
Properties: Model training was conducted on text datasets using question–passage pairs from publicly available, commercially permissible datasets and synthetically generated datasets.
Testing Dataset:
Data Collection Method by dataset: Not Applicable
Labeling Method by dataset: Not Applicable
Properties: Not Applicable. Model quality was assessed using the evaluation benchmark datasets described in the Evaluation Dataset subsection.
Evaluation Dataset:
Data Collection Method by dataset: Hybrid: Human, Automated, Synthetic
Labeling Method by dataset: Hybrid: Human, Automated, Synthetic
Properties: This model is evaluated on 16 public tasks on Retrieval Embedding Benchmark (RTEB), a new benchmark designed to reliably evaluate the retrieval accuracy of embedding models for real-world applications. More details on RTEB can be found on their leaderboard.
We also evaluated the model on the MMTEB (Retrieval) benchmark datasets (paper), and on the eight text datasets (extracted via OCR) from ViDoRe-V3 benchmark.
We set the model sequence length to 4096 for the evaluation results below.
Inference:
Acceleration Engine: PyTorch, vLLM
Test Hardware: NVIDIA Ampere - A100 80GB NVIDIA Hopper - H100 80GB
Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.
Please report model quality, risk, security vulnerabilities or NVIDIA AI concerns here.
Bias
Table with columns: Field, Response| Field | Response |
|---|
| Participation considerations from adversely impacted groups protected classes in model design and testing: | None |
| Measures taken to mitigate against unwanted bias: | None |
| Bias Metric (If Measured): | None |
Explainability
Table with columns: Field, Response| Field | Response |
|---|
| Intended Task/Domain: | Passage and query embedding for question and answer retrieval |
| Model Type: | Transformer encoder |
| Intended Users: | Generative AI creators working with conversational AI models - users who want to build a multilingual question and answer application over a large text corpus, leveraging the latest dense retrieval technologies. |
| Output: | Array of float numbers (Dense Vector Representation for the input text) |
| Describe how the model works: | Model transforms the tokenized input text into a dense vector representation. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable |
| Technical Limitations & Mitigation: | The model’s max sequence length is 32768. Therefore, the longer text inputs should be truncated. |
| Verified to have met prescribed NVIDIA quality standards: | Yes |
| Performance Metrics: | Accuracy, Throughput, and Latency |
| Potential Known Risks: | This model does not always guarantee to retrieve the correct passage(s) for a given query. |
| Licensing: | This model and its associated configuration files are licensed under the OpenMDW License Agreement, version 1.1 (OpenMDW-1.1). Additional Information: Built with Ministral-3-3B-Instruct-2512 which is released under Apache 2.0. |
Privacy
Table with columns: Field, Response| Field | Response |
|---|
| Generatable or reverse engineerable personal data? | None |
| Was consent obtained for any personal data used? | Not Applicable |
| Personal data used to create this model? | None Known |
| How often is the dataset reviewed? | Before Every Release |
| Is there provenance for all datasets used in training? | Yes |
| Does data labeling (annotation, metadata) comply with privacy laws? | Yes |
| Was data from user interactions with the AI model (e.g. user input and prompts) used to train the model? | Yes |
| Is data compliant with data subject requests for data correction or removal, if such a request was made? | No, not possible with externally-sourced data. |
| Applicable Privacy Policy | https://www.nvidia.com/en-us/about-nvidia/privacy-policy/ |
Safety
Table with columns: Field, Response| Field | Response |
|---|
| Model Application(s): | Text Embedding for Retrieval |
| Describe the physical safety impact (if present). | Not Applicable |
| Use Case Restrictions: | This model and its associated configuration files are licensed under the OpenMDW License Agreement, version 1.1 (OpenMDW-1.1). Additional Information: Built with Ministral-3-3B-Instruct-2512 which is released under Apache 2.0. |
| Model and dataset restrictions: | The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to. |
