NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8

NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8

<div align="center"> <img src="./star_elastic_logo.png" alt="Nemotron Labs Elastic" width="500"/> </div> <div align="center" style="line-height: 1;"> <a href="https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8/resolve/main/star_elastic_arxiv.pdf" target="_blank" style="margin: 2px;"> <img alt="Paper" src="https://img.shields.io/badge/📝Paper-Read Now!-536af5?color=76B900&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://icml.cc/virtual/2026/poster/61782" target="_blank" style="margin: 2px;"> <img alt="ICML 2026" src="https://img.shields.io/badge/ICML_2026-Poster-blue?color=0072C6&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://huggingface.co/collections/nvidia/nemotron-pre-training-datasets" target="_blank" style="margin: 2px;"> <img alt="Pre-Training Datasets" src="https://img.shields.io/badge/🗄️_Pre--Training_Datasets-Available_Here-76B900?logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://huggingface.co/collections/nvidia/nemotron-post-training-v3" target="_blank" style="margin: 2px;"> <img alt="Post-Training Datasets" src="https://img.shields.io/badge/🗄️_Post--Training_Datasets-Available_Here-76B900?logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> </div> <div align="center" style="line-height: 1;"> <a href="https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/" style="margin: 2px;"> <img alt="License" src="https://img.shields.io/badge/License-NVIDIA Open Model License-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/> </a> </div>

Model Developer: NVIDIA

Model Dates:

September 2025 - December 2025

Data Freshness:

• The post-training data has a cutoff date of November 28, 2025.
• The pre-training data has a cutoff date of June 25, 2025.

Model Overview

NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8 is a 3-in-1 elastic large language model (LLM) developed by NVIDIA. It contains three nested model variants (30B, 23B, and 12B parameters) within a single FP8 checkpoint, all sharing the same parameter space. The 23B and 12B variants can be extracted zero-shot from this checkpoint using the provided slicing script.

This is the FP8 quantized version of the BF16 elastic model. Weights are stored as float8_e4m3fn with per-tensor weight_scale and input_scale scalars. The FP8 quantization preserves the nested weight-sharing structure -- scale tensors are per-tensor scalars and are not pruned during slicing, while actual weight tensors are pruned along the appropriate axes.

This model was derived from NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 using the Elastic post-training framework, which produces elastic (many-in-one) reasoning LLMs from hybrid Mamba-Transformer-MoE architectures. The method takes the parent reasoning model and embeds N nested submodels within it using the compute of a single training run. The technology was originally developed for hybrid Mamba-Transformer architectures and has been extended to hybrid MoE models for Nemotron 3 Nano.

📄 The accompanying paper, "Star Elastic: Many-in-One Reasoning LLMs with Efficient Budget Control", has been accepted for publication at ICML 2026 (see Citation).

The parent model employs a hybrid Mixture-of-Experts (MoE) architecture, consisting of 23 Mamba-2 and MoE layers, along with 6 Attention layers. Each MoE layer includes 128 experts plus 1 shared expert, with 6 experts activated per token.

All three nested variants share the same 52-layer architecture pattern and the same number of attention heads (32), Mamba heads (64), and MoE experts (128). The variants differ in embedding dimension and MoE FFN dimension:

The supported languages include: English, German, Spanish, French, Italian, and Japanese.

This model is ready for commercial use.

Average accuracy of Elastic variants compared to the parent Nemotron 3 Nano 30B and Qwen3-30B-A3B across key reasoning benchmarks (BF16 precision). The Elastic-30B variant matches or exceeds the parent model on most benchmarks, while the 23B and 12B variants provide strong accuracy at reduced compute.

Quantized Accuracy Recovery

The following table summarizes accuracy recovery of quantized variants relative to BF16:

Throughput Improvements

The smaller elastic variants provide substantial throughput improvements when served with vLLM (measured on H100 GPU, ISL=8192 / OSL=16384, BF16 precision):

The smaller nested models also enable much higher batch sizes on the same GPU (224 vs 36), providing significant serving cost reductions.

Elastic Budget Control

Elastic Budget Control is a novel inference-time mechanism enabled by the nested architecture. Instead of using a fixed model for both the thinking (<think>) and answering phases, elastic budget control uses different-sized nested models for each phase.

Pareto frontier of accuracy vs. latency for different elastic budget control configurations. The figure shows that using different model sizes for the thinking and answering phases (e.g., 23B thinking -> 30B answering) achieves better accuracy-latency tradeoffs than using a single model size throughout. The optimal configuration uses a smaller model for high-volume reasoning (thinking phase) and a larger model for high-fidelity synthesis (answering phase), achieving up to 16% higher accuracy and 1.9x lower latency compared to standard single-model budget control.

Four configurations are possible using models M_L (Large) and M_S (Small):

• M_L -> M_L: Large model for both thinking and answering
• M_S -> M_S: Small model for both thinking and answering
• M_L -> M_S: Large model for thinking, small for answering
• M_S -> M_L: Small model for thinking, large for answering (identified as optimal)

The M_S -> M_L configuration is optimal because:

1. Thinking phase (high-volume reasoning): Benefits from larger token budgets to explore reasoning paths; using the smaller model generates extensive reasoning traces with minimal computational overhead.
2. Answering phase (high-fidelity synthesis): Requires superior instruction-following and consistency; the larger model provides the necessary capacity for robust synthesis.

The 23B -> 30B configuration achieves the best accuracy-latency tradeoffs over a wide range of budgets.

⚠️ Note on inference support. Elastic budget control is not yet supported in the standard vLLM inference engine — switching nested sub-models within a single generation (e.g., 23B → 30B think → answer) currently requires a custom inference path. Nested models preserve the Mamba and attention layer structure, enabling cache-state transplantation between models, and efficient native vLLM integration is actively being worked on.

Elastic Architecture Overview

A key innovation of this model is its Elastic Architecture, which enables the extraction of smaller, nested variants (23B and 12B parameters) from the same parameter space without requiring separate training runs.

The Elastic pipeline operates in three stages:

1. Importance Estimation: Components (embedding dimensions, attention heads, Mamba heads, MoE experts, FFN channels) are ranked by importance scores using calibration data.
2. Elastic Formulation: Smaller-budget sub-networks are defined as contiguous subsets of the most salient components, forming a nested hierarchy.
3. Elastic Training: A learnable router with Gumbel-Softmax selection is trained end-to-end using knowledge distillation from the frozen parent model, with a two-stage curriculum (8K context, then 49K context).

The entire elastic family was produced by post-training the Nemotron 3 Nano 30B parent with only approximately 160B tokens — roughly 0.6% of the parent's ~25T-token pretraining budget, and far less than what would be required to train three independent compressed variants.

Key Benefits

• 3-in-1 Nested Checkpoint: All three model sizes (12B/23B/30B) are embedded in a single checkpoint via nested weight sharing. Deploying all three variants requires only 58.9 GB in BF16, a 2.14x memory reduction compared to storing three independent checkpoints (126.1 GB). The FP8 checkpoint provides further memory reduction.
• Compute-efficient elastification: The entire 30B + 23B + 12B nested family was produced by post-training the Nemotron 3 Nano 30B parent in a single training run with only ~160B tokens — about 0.6% of the parent's ~25T pretraining budget, and far less than retraining or independently compressing three separate models.
• Full Accuracy-Latency Frontier: The nested model covers all latency regimes via elastic budget control, delivering up to 16% higher accuracy and 1.9x faster inference compared to standard single-model budget control.
• High FP8 Accuracy Recovery: FP8 quantization preserves strong accuracy across all variants (see Quantized Accuracy Recovery below).
• Unlocks lower-tier RTX community prototyping: As a bonus, the 12B and 23B variants in FP8 / NVFP4 fit on consumer/prosumer RTX series lower tiers, including 6000 / 5090 / 5080. This should make Nano V3 architecture easier to prototype and debug.

Zero-Shot Slicing Before Deployment

This checkpoint contains the full 30B 3-in-1 FP8 model. Before deploying a smaller variant, you can perform zero-shot slicing to extract the 23B or 12B model directly from this checkpoint — no additional training or fine-tuning is required. The sliced model is ready for immediate deployment.

Use the provided zero_shot_slicing.py script:

# Zero-shot slice the 23B FP8 variant for deployment
python zero_shot_slicing.py \
    --source-checkpoint <path-to-this-30B-fp8-checkpoint> \
    --target-checkpoint ./nemotron-elastic-23b-fp8 \
    --size 23B \
    --precision fp8

# Zero-shot slice the 12B FP8 variant for deployment
python zero_shot_slicing.py \
    --source-checkpoint <path-to-this-30B-fp8-checkpoint> \
    --target-checkpoint ./nemotron-elastic-12b-fp8 \
    --size 12B \
    --precision fp8

The zero-shot slicing process preserves the hybrid MoE architecture while reducing model size through structured pruning of embedding dimensions and MoE FFN dimensions. Because the nested variants share the most salient weights with the parent model, the sliced checkpoints retain strong accuracy without any additional knowledge distillation or fine-tuning. FP8 scale tensors (weight_scale, input_scale) are per-tensor scalars and are left unchanged during slicing. The shared expert intermediate size (3712) is also left unchanged.

Available size presets:

Deployment Memory

Quick Start Guide

Use it with Transformers

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load the full 30B FP8 elastic model
tokenizer = AutoTokenizer.from_pretrained("nvidia/NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "nvidia/NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Write a haiku about GPUs"},
]

tokenized_chat = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    tokenized_chat,
    max_new_tokens=1024,
    temperature=1.0,
    top_p=1.0,
    eos_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0]))

temperature=1.0 and top_p=1.0 are recommended for reasoning tasks.

If you'd like to use reasoning off, add enable_thinking=False to apply_chat_template(). By default, enable_thinking is set to be True.

Note: To use the 23B or 12B FP8 variants, first extract them using the slicing script as described in the Zero-Shot Slicing Before Deployment section, then load the extracted checkpoint.

Use it with vLLM

For more detailed information on how to use the model with vLLM, please see this cookbook. If you are on Jetson Thor or DGX Spark, please use this vllm container.

pip install -U "vllm>=0.12.0"

Download the custom parser from the Hugging Face repository.

wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py

Launch a vLLM server using the custom parser.

vllm serve nvidia/NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8 \
  --served-model-name model \
  --max-num-seqs 8 \
  --tensor-parallel-size 1 \
  --max-model-len 131072 \
  --port 8000 \
  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser-plugin nano_v3_reasoning_parser.py \
  --reasoning-parser nano_v3

In the example above, we use a context length of 128k. You can increase the context size up to 1M to support longer contexts. To enable this, set the VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 environment variable as shown below:

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
vllm serve nvidia/NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8 \
  --served-model-name model \
  --max-num-seqs 8 \
  --tensor-parallel-size 1 \
  --max-model-len 1M \
  --port 8000 \
  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser-plugin nano_v3_reasoning_parser.py \
  --reasoning-parser nano_v3

Here is an example client code for vLLM. By default, the endpoint has reasoning enabled. We recommend setting a high value (e.g., 10,000) for max_tokens.

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "model",
        "messages":[{"role": "user", "content": "Write a haiku about GPUs"}],
        "max_tokens": 10000
    }'

If you'd like to use reasoning off with vLLM, you can do the following:

vLLM OpenAI curl request:

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "model",
        "messages":[{"role": "user", "content": "Write a haiku about GPUs"}],
        "chat_template_kwargs": {"enable_thinking": false}
    }'

vLLM OpenAI client:

response = client.chat.completions.create(model=model, messages=messages, extra_body={"chat_template_kwargs": {"enable_thinking": False}})

Note: To serve the 23B or 12B variants with vLLM, first extract them using the zero-shot slicing script as described in the Zero-Shot Slicing Before Deployment section, then point vLLM to the sliced checkpoint.

Prompt Format

The prompt format is the same as the NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model. Please refer to that model card for detailed prompt formatting information and usage examples.

License/Terms of Use

Your use of this model is governed by the NVIDIA Open Model License.

Deployment Geography

Global

Use Case

This model is intended for developers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. Supported for English and coding languages, with additional support for Spanish, French, German, Japanese, and Italian.

Release Date

• Hugging Face: 05/07/2026 via NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8

Model Architecture

• Architecture Type: Mamba2-Transformer Hybrid Mixture of Experts (MoE)
• Network Architecture: Nemotron Hybrid MoE
• Number of model parameters: 30B (full), 23B and 12B (nested variants)
• Precision: FP8 (float8_e4m3fn with per-tensor scales)
• Number of layers: 52
• Layer pattern: M-E-M-E-M*-E-M-E-M-E-M*-E-... (M=Mamba, E=MoE, *=Attention)
• Number of attention heads: 32
• Number of Mamba heads: 64
• Number of MoE experts: 128 routed + 1 shared (6 active per token)

Input

• Input Type(s): Text
• Input Format(s): String
• Input Parameters: One-Dimensional (1D): Sequences
• Maximum input size: 1M tokens
• Other Properties Related to Input: Supported languages include: English, Spanish, French, German, Japanese, Italian

Output

• Output Type(s): Text
• Output Format: String
• Output Parameters: One-Dimensional (1D): Sequences
• Maximum output size: 128K tokens
• Other Properties Related to Output: Context length up to 128K tokens for output generation. Text-only output in supported languages (English, Spanish, French, German, Japanese, Italian).

Our AI models are designed and 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.

Software Integration

• Runtime Engine(s): Hugging Face Transformers
• Supported Hardware Microarchitecture Compatibility: NVIDIA Hopper, NVIDIA Blackwell
• Supported Operating System(s): Linux

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.

This AI model can be embedded as an Application Programming Interface (API) call into the software environment described above.

Model Version(s)

• NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-FP8 v1.0 — Elastic 3-in-1 FP8 quantized model containing 30B, 23B, and 12B nested variants (quantized, deployable)

Training Methodology

This elastic model was post-trained from NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 using the Elastic framework with knowledge distillation.

Elastic Training Details:

• Total training budget: ~160B tokens
• Stage 1: Uniform budget sampling, sequence length 8192, ~100B tokens
• Stage 2: Curriculum-based non-uniform sampling (skewed toward larger budgets), sequence length 49152, ~60B tokens
• Optimizer: LR 1e-4 for model, LR 1e-2 for router, 60-step linear warmup
• Data blend: 70% reasoning (post-training/SFT) + 30% pretraining data
• Loss: KL divergence between frozen parent (teacher) and elastic model (student) + router resource cost loss

The extended context in Stage 2 is critical for reasoning performance. The two-stage curriculum (8K then 49K context) outperforms training at 49K from the start.

The FP8 quantization was applied after elastic training.

For details on the parent model's pre-training, supervised fine-tuning, and reinforcement learning stages, please refer to the NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model card.

Training and Evaluation Datasets

The elastic post-training used a blend of 70% reasoning (post-training/SFT) data and 30% pretraining data from the parent model's training corpus. The full dataset disclosure below is inherited from the parent model NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.

Data Modality: Text
The total size: 10,648,823,153,919 Tokens
Total number of datasets: 141
Dataset partition: Training [100%], testing [0%], validation [0%]
Time period for training data collection: 2013 to May 1, 2025
Time period for testing data collection: 2013 to May 1, 2025
Time period for validation data collection: 2013 to May 1, 2025
Data Collection Method by dataset: Hybrid: Automated, Human, Synthetic
Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 is pre-trained on a large corpus of high-quality curated and synthetically-generated data. It is trained in the English language, as well as 19 other languages and 43 programming languages. Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials. The corpus spans domains including legal, math, science, finance, and more. We also include a small portion of question-answering, and alignment style data to improve model accuracy. The model was trained for approximately 25 trillion tokens.

The post-training corpus for NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 of high-quality curated and synthetically-generated data. Primary languages used for post-training include English, German, Spanish, French, Italian, and Japanese.

These datasets, such as FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3, do not collectively or exhaustively represent all demographic groups (and proportionally therein). For instance, these datasets do not contain explicit mentions of demographic classes such as age, gender, or ethnicity in 64-99% of samples, depending on the source. In the subset where such terms are present, document-based datasets (FinePDFs and EssentialWeb) contain representational skews, such as references to "male" outnumbering those to "female", and mentions of "White" as the most frequent among ethnic identifiers (comprising 43-44% of ethnicity mentions). To mitigate these imbalances, we recommend considering evaluation techniques such as bias audits, fine-tuning with demographically balanced datasets, and mitigation strategies like counterfactual data augmentation to align with the desired model behavior. This evaluation used a 3,000-sample subset per dataset, identified as the optimal threshold for maximizing embedder accuracy.

During post-training, we generate synthetic data by distilling trajectories, solutions, and translations from strong teacher models and agent systems, often grounded in real tasks or documents and aggressively filtered for quality. For math, code, and science, we start from curated problem sets and use open source permissive models such as GPT-OSS-120B to produce step-by-step reasoning traces, candidate solutions, best-of-n selection traces, and verified CUDA kernels. For long-context and science, we build synthetic QA and reasoning data by retrieving passages from long documents, generating MCQ/OpenQA questions and answers, and paraphrasing them into multiple prompt/response formats to ensure diversity. Across all pipelines we stack automated verification—compilers, numerical checks, language identification—to ensure our data is high quality.

For all domains, we apply a unified data filtering pipeline to ensure that only high-quality, license-compliant, and verifiable samples are used for post-training. We first discard malformed examples using structural checks (e.g., missing tool definitions when tool calls are present). We then aggressively filter reasoning traces exhibiting pathological repetition, such as repeated n-grams within a sliding window or across the entire trajectory, which we found to be a strong indicator of malformed or low-quality reasoning. Finally, based on internal audits of synthetically generated datasets, we observed that some teacher models occasionally produce reasoning traces and final responses that implicitly align with specific political entities or promote nationalistic narratives. To mitigate this, we apply targeted keyword- and regex-based filters and remove all trajectories matching such behavior.

Alongside the model, we release our final pre-training and post-training data, as outlined in this section. For ease of analysis, there is a sample set that is ungated. For all remaining code, math and multilingual data, gating and approval is required, and the dataset is permissively licensed for model training purposes.

More details on the datasets and synthetic data generation methods can be found in the technical report NVIDIA Nemotron 3 Nano.

Private Non-publicly Accessible Datasets of Third Parties

Private Non-publicly Accessible Datasets by NVIDIA

Crawled and Scraped from Online Sources by NVIDIA

The English Common Crawl data was downloaded from the Common Crawl Foundation (see their FAQ for details on their crawling) and includes the snapshots CC-MAIN-2013-20 through CC-MAIN-2025-13. The data was subsequently deduplicated and filtered in various ways described in the Nemotron-CC paper. Additionally, we extracted data for fifteen languages from the following three Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The fifteen languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not have reliable multilingual model-based quality classifiers available, we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline, but selectively removing some filters for some languages that did not work well. Deduplication was done in the same way as for Nemotron-CC.

The GitHub Crawl was collected using the GitHub REST API and the Amazon S3 API. Each crawl was operated in accordance with the rate limits set by its respective source, either GitHub or S3. We collect raw source code and subsequently remove any having a license which does not exist in our permissive-license set (for additional details, refer to the technical report).

NVIDIA-Sourced Synthetic Datasets

Training Dataset

We use a considerable amount of synthetic data. Out of 10.6 trillion tokens, 3,534,013,958,278 tokens are synthetically generated.

We extracted data for fifteen languages from the following three Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The fifteen languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not have reliable multilingual model-based quality classifiers available, we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline, but selectively removing some filters for some languages that did not work well. Deduplication was done in the same way as for Nemotron-CC. Additionally, we used data from Wikipedia and FineWeb-2 (Penedo et al., 2025) for these fifteen languages as well as four additional languages: Czech, Finnish, Hebrew, and Hindi.

We collect a total of 922,476,782,017 tokens of code in 43 different languages.

Language Distribution in Post-Training

For our post-training recipe, we focused on 5 main languages in addition to English: Spanish, French, Japanese, Italian, German.
Those languages were represented in the form of multilingual reasoning and translation task.

The following table depicts our sample distribution for the 6 languages and 5 translation pairs.

Evaluation Dataset

• Data Collection Method by dataset: Hybrid: Human, Synthetic
• Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

Inference

• Acceleration Engine: HF, vLLM, TRT-LLM, SGLang, Llama.cpp
• Test Hardware: NVIDIA A100 80GB, H100 80GB, B200 192GB, RTX PRO 6000 96GB, Jetson Thor, DGX Spark

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. When downloaded or used in accordance with our Trustworthy AI terms of service, 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++ subcards: Bias, Explainability, Privacy, and Safety.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Benchmark Results

Reasoning Evaluations (BF16 Precision)

Key observations: Elastic-30B matches or exceeds the parent NanoV3-30B on most benchmarks (e.g., AIME-2025: 88.54 vs 87.92, IFBench instruct: 73.96 vs 73.19, Tau-Retail: 59.36 vs 53.51). Both Elastic-23B and Elastic-12B outperform Qwen3-30B-A3B on AIME-2025 and IFBench by large margins.

Citation

If you use this model, please cite our paper, accepted to ICML 2026:

@inproceedings{taghibakhshi2026starelastic,
  title     = {Star Elastic: Many-in-One Reasoning {LLMs} with Efficient Budget Control},
  author    = {Taghibakhshi, Ali and Cai, Ruisi and Muralidharan, Saurav and Turuvekere Sreenivas, Sharath and Mahabaleshwarkar, Ameya and Chochowski, Marcin and Bercovich, Akhiad and Zilberstein, Ran and El-Yaniv, Ran and Geifman, Yonatan and Korzekwa, Daniel and Suhara, Yoshi and Olabiyi, Oluwatobi and Aithal, Ashwath and Tajbakhsh, Nima and Molchanov, Pavlo},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  series    = {ICML 2026},
  year      = {2026},
  note      = {Accepted}
}

@article{taghibakhshi2025nemotronelastic,
  title   = {Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs},
  author  = {Taghibakhshi, Ali and Turuvekere Sreenivas, Sharath and Muralidharan, Saurav and Cai, Ruisi and Chochowski, Marcin and Mahabaleshwarkar, Ameya Sunil and Suhara, Yoshi and Olabiyi, Oluwatobi and Korzekwa, Daniel and Patwary, Mostofa and Shoeybi, Mohammad and Kautz, Jan and Catanzaro, Bryan and Aithal, Ashwath and Tajbakhsh, Nima and Molchanov, Pavlo},
  journal = {arXiv preprint arXiv:2511.16664},
  year    = {2025},
  url     = {https://arxiv.org/abs/2511.16664}
}

@article{blakeman2025nemotron,
  title   = {Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning},
  author  = {Blakeman, Aaron and Grattafiori, Aaron and Basant, Aarti and Gupta, Abhibha and Khattar, Abhinav and Renduchintala, Adi and Vavre, Aditya and Shukla, Akanksha and Bercovich, Akhiad and Ficek, Aleksander and others},
  journal = {arXiv preprint arXiv:2512.20848},
  year    = {2025},
  url     = {https://arxiv.org/abs/2512.20848}
}