Nemotron-3-Embed-1B-NVFP4

Nemotron-3-Embed-1B-NVFP4

Nemotron-3-Embed-1B-NVFP4 is the quantized version of the Nemotron-3-Embed-1B-BF16 model, developed for text question-answering retrieval. The model is quantized with NVIDIA Model Optimizer, using nvidia-modelopt v0.45.0.

Evaluated on 34 languages: English, Arabic, Assamese, Bengali, Bulgarian, Chinese, Danish, Dutch, Finnish, French, German, Hindi, Hinglish, Indonesian, Italian, Japanese, Korean, Malay, Marathi, Nepali, Norwegian, Persian, Portuguese, Romanian, Russian, Spanish, Swahili, Swedish, Tamil, Telugu, Thai, Ukrainian, Urdu, Vietnamese.

This model is ready for commercial use.

License / Terms of Use

Licensed under the OpenMDW License Agreement, version 1.1 (OpenMDW-1.1).

Additional Information: Built with Ministral-3-3B-Instruct-2512, released under Apache 2.0.

This project will download and install additional third-party open source software projects. Review the license terms of these open source projects before use.

Deployment Geography

Global

Use Case

The Nemotron-3-Embed-1B-NVFP4 is most suitable for users who want to build a multilingual question-and-answer application over a large text corpus, leveraging the latest dense retrieval technologies.

Generally, Nemotron-3-Embed-1B-BF16 and Nemotron-3-Embed-1B-NVFP4 models share the embedding space and can be used interchangeably, but validating retrieval quality on a representative sample before switching models is recommended.

Release Date

07/16/2026 via huggingface.co/nvidia/Nemotron-3-Embed-1B-NVFP4

Model Architecture

Nemotron-3-Embed-1B-NVFP4 is the quantized version of Nemotron-3-Embed-1B-BF16, 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.

Input / Output

Input

• Type: Text
• Format: List of strings
• Parameters: One-Dimensional (1D)
• Notes: Text inputs should be tokenized by the model tokenizer. Max sequence length is 32768. Longer inputs should be chunked or truncated.

Output

• Type: Floats
• Format: List of float arrays
• Parameters: One-Dimensional (1D) embedding vector per input text string
• Notes: Outputs a 2048-dimensional embedding vector per input text string. Supports dynamic embedding sizes by slicing the vector from the start (e.g., keeping the first 1024 or 512 dimensions). Sliced embeddings remain highly functional, provided the resulting sub-vector is re-normalized (L2 normalization) after slicing.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems, leveraging NVIDIA's hardware (GPU cores) and software frameworks (CUDA libraries) for faster training and inference compared to CPU-only solutions.

Post-Training Quantization and Quantization-Aware Distillation

This model is a post-training quantized variant of nvidia/Nemotron-3-Embed-1B-BF16. Quantization was applied to the weights and activations of linear layers only, targeting the NVFP4 data type for efficient inference. Quantization-Aware Distillation (QAD) was applied primarily to recover accuracy for long input sequences.

vLLM Usage

This NVFP4 checkpoint is intended for vLLM. The examples below show offline Python use and online serving. Use nvidia/Nemotron-3-Embed-1B-BF16 if you need Transformers or Sentence Transformers.

Use the Hugging Face model ID by default. If working from a local checkpoint, replace MODEL_ID with that path:

MODEL_ID=nvidia/Nemotron-3-Embed-1B-NVFP4

The output tables use q[i] for queries and d[i] for documents. Scores are rounded to four decimal places, and runtime differences might affect the final decimal places.

Tested vLLM Versions

This checkpoint has been explicitly tested with the following vLLM distributions:

Use vLLM 0.25.0 for the examples below. vLLM 0.23.x and 0.24.x have known issues with this NVFP4 checkpoint family. Other versions have not been explicitly validated.

To install vLLM without a container image:

pip install --upgrade "vllm==0.25.0" openai requests numpy

vLLM Offline Python

Use the offline Python API for local vLLM inference without an HTTP server. LLM.embed accepts formatted strings — add the query: and passage: prefixes manually. The example uses a 4,096-token limit. Review CUDA graph sizing before increasing it.

Expected output:

Similarity scores:
          d[0]      d[1]      d[2]      d[3]
q[0]    0.8064   0.0201   0.0003  -0.0320
q[1]    0.0445   0.6469  -0.0516   0.0388
q[2]   -0.0083  -0.0402   0.6558   0.1071
q[3]   -0.0222   0.0265   0.1261   0.7677

vLLM Online Serving

Tested vLLM builds read the checkpoint NVFP4 metadata, so no quantization flag is required.

MODEL_ID=nvidia/Nemotron-3-Embed-1B-NVFP4
MAX_MODEL_LEN=4096
MAX_BATCHED_TOKENS=4096

vllm serve "$MODEL_ID" \
  --max-model-len "$MAX_MODEL_LEN" \
  --max-num-batched-tokens "$MAX_BATCHED_TOKENS" \
  --max-cudagraph-capture-size "$MAX_BATCHED_TOKENS"

CUDA Graph Sizing

The checkpoint supports sequences up to 32,768 tokens. The examples use 4,096 as a conservative starting point.

Guidance for tuning CUDA graph capture:

• Set --max-model-len to the longest request you intend to serve. Tune --max-num-batched-tokens for the workload, concurrency, and available GPU memory. When chunked prefill is disabled, the batched-token budget must be at least the model-length limit.
• For default capture buckets up to 8,192, set --max-cudagraph-capture-size equal to --max-num-batched-tokens. This makes batches up to the scheduler budget eligible for CUDA graph execution. Batches outside the captured range use a slower uncaptured path.
• Larger capture ranges increase startup time and graph memory. Benchmark representative traffic on the target hardware. Do not capture beyond the batched-token budget.
• Use a maximum capture size of 4,096 as the conservative default for services that restart or autoscale regularly. A maximum capture size of 8,192 can be reasonable when a longer cold start is acceptable. For maximum capture sizes above 8,192, pass a smaller, workload-aligned set with --cudagraph-capture-sizes to keep startup time under control.

Example Sparse Capture Sizes for Inputs up to 32,768 Tokens

MODEL_ID=nvidia/Nemotron-3-Embed-1B-NVFP4
MAX_BATCHED_TOKENS=32768

vllm serve "$MODEL_ID" \
  --max-num-batched-tokens "$MAX_BATCHED_TOKENS" \
  --cudagraph-capture-sizes \
    1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128 \
    136 144 152 160 168 176 184 192 200 208 216 224 232 240 248 256 \
    384 512 768 1024 1536 2048 3072 4096 6144 8192 12288 16384 \
    24576 32768

Choose sizes from representative batch-token measurements. vLLM pads each execution batch to the next captured size, so denser lists reduce padding but require more startup time and graph memory.

In illustrative tests on an NVIDIA GB10 system with vLLM 0.25.0, cold startup using automatic buckets took about 74 seconds at a maximum capture size of 4,096 and 121 seconds at 8,192. At 32,768, startup with automatic buckets was projected to take tens of minutes. With the sparse list of 49 capture sizes above, startup completed in about one minute. Results vary by workload and hardware.

Add --host or --port to the serving command for non-default network settings.

To serve a local checkpoint, replace MODEL_ID with its path. Add --served-model-name nvidia/Nemotron-3-Embed-1B-NVFP4 if clients should continue using the Hugging Face model ID.

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 prompt metadata.

import numpy as np
import requests

MODEL = "nvidia/Nemotron-3-Embed-1B-NVFP4"
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))

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:.

Expected Configuration Warning

When vLLM loads this checkpoint, its Transformers configuration parser can emit the following warning:

[transformers] Unrecognized keys in `rope_parameters` for 'rope_type'='yarn': {'apply_yarn_scaling'}

This warning is expected and does not prevent vLLM from loading the model or running inference. apply_yarn_scaling is a temporary vLLM compatibility field that preserves the checkpoint's intended long-context rotary position embedding (RoPE) behavior. Do not remove it from config.json. Refer to vLLM issue #48621 for upstream compatibility work.

Software Integration

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-NVFP4

Training, Testing, and Evaluation Datasets

Dataset Overview

This checkpoint is an NVFP4 post-training-quantized derivative of nvidia/Nemotron-3-Embed-1B-BF16. Quantization-Aware Distillation (QAD) was applied on a small sample of the original BF16 data mix for further accuracy recovery.

Calibration Dataset

• Data Collection Method by dataset: Automated
• Labeling Method by dataset: Automated
• Properties: A calibration dataset of 512 total samples was used for NVFP4 post-training quantization, consisting of 256 queries and 256 passages from the abisee/cnn_dailymail dataset, formatted with query and passage prefixes.

Training Dataset

QAD training was performed as part of the NVFP4 model development. The information below describes the QAD training datasets used for this model. For details about the training datasets used to build the underlying base model, Nemotron-3-Embed-1B-BF16, refer to its model card.

• Total Size: 20k data samples
• Total Number of Dataset Files: 5
• Dataset Partition: Training [100%], Testing [N/A — evaluation benchmarks used separately], Validation [N/A — evaluation benchmarks used separately]

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.

LLMs used to generate synthetic datasets:

• nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
• nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4

Seed Datasets:

Data Modality: Text

Training Data Size: Text Training Data Size: 20k

• Data Collection Method by dataset: Hybrid: Human, Automated, Synthetic
• Labeling Method by dataset: Hybrid: Human, Automated, Synthetic
• Properties: Model distillation training of the underlying model was conducted on text datasets using question–passage pairs from publicly available, commercially permissible datasets and synthetically generated datasets. For more information, refer to the base model card.

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 section compares the performance of the quantized model Nemotron-3-Embed-1B-NVFP4 with the baseline implementation Nemotron-3-Embed-1B-BF16.

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.

Model sequence length was set to 4096 for the evaluation results below. The NVFP4 model was evaluated on an NVIDIA GB200 GPU.

Text Retrieval Benchmarks (chunk retrieval) — Avg. NDCG@10

Inference

• Acceleration Engine: vLLM
• Test Hardware:
  • NVIDIA Ampere - A100 PCIe/SXM
  • NVIDIA Blackwell - GB200 and RTX 6000 PRO
  • NVIDIA Hopper - H100 PCIe/SXM
  • NVIDIA Lovelace - L40 and L4

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility, and has 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, see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.

Report model quality, risk, security vulnerabilities, or NVIDIA AI concerns here.

Bias

Explainability

Privacy

Safety