nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 · Hugging Face

Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#descriptionDescription:

Nemotron-Labs-3-Puzzle-75B-A9B is a deployment-optimized large language model developed by NVIDIA, derived from Nemotron-3-Super-120B-A12B. The model is produced using Iterative Puzzle, a post-training compression framework, with the goal of significantly improving inference efficiency for interactive, reasoning-heavy, and long-context workloads while preserving strong downstream accuracy.

The model employs a hybrid MoE architecture with interleaved Mamba, MoE, and Attention layers. Like Nemotron-3-Super, it supports Multi-Token Prediction (MTP) for faster text generation. Compared to its parent, Puzzle-75B-A9B reduces the model from 120.7B total / 12.8B active parameters to 75.3B total / 9.3B active parameters.

See the tech report for full training and compression details:Nemotron-Labs-3-Puzzle-75B-A9B: Compressing Hybrid MoE LLMs.

Compared to Nemotron-3-Super, Puzzle-75B-A9B:

• Achieves approximately 2× higher server throughput on a single 8×B200 node at matched user-throughput constraints,
• Increases sustainable 1M-token single-H100 concurrency from 1 request to 8 requests,
• Maintains strong accuracy across reasoning, coding, multilingual, long-context, and agentic benchmarks.

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

This model is ready for commercial use.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#licenseterms-of-useLicense/Terms of Use

**Governing Download Terms:**Use of this model is governed by theOpenMDW License Agreement, version 1.1(OpenMDW-1.1).

This project is currently not accepting contributions.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#benchmarksBenchmarks

BenchmarkNemotron-Labs-3-Puzzle-75B-A9B-BF16Nemotron-Labs-3-Puzzle-75B-A9B-FP8Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4General KnowledgeMMLU-Pro82.482.082.2ReasoningAIME25 (no tools)89.789.489.9HMMT Feb25 (no tools)93.492.792.9HMMT Feb25 (with tools)93.993.693.1GPQA (no tools)78.677.878.0GPQA (with tools)79.580.678.2LiveCodeBench (v5 2024-07↔2024-12)81.180.579.9SciCode (subtask)40.639.640.3HLE (no tools)16.516.015.7AgenticTerminal Bench (hard subset)24.022.923.4TauBench V2Airline55.854.555.7Retail63.263.463.7Telecom61.561.360.3Average60.259.759.9Chat & Instruction FollowingIFBench (prompt)71.971.971.3Scale AI Multi-Challenge56.655.455.9Arena-Hard-V268.669.869.0Long ContextAA-LCR56.956.657.1RULER @ 256k95.195.395.3RULER @ 512k94.294.594.8RULER @ 1M92.292.493.2MultilingualMMLU-ProX (avg over langs)77.577.176.5WMT24++ (en→xx)85.285.285.1 All evaluation results were collected viaNemo Evaluator SDKand for most benchmarks, theNemo Skills Harness. For reproducibility purposes, more details on the evaluation settings can be found in theNemo Evaluator SDK configs folderand thereproducibility tutorial for Nemotron 3 Super. The open source container on Nemo Skills packaged via NVIDIA’s Nemo Evaluator SDK used for evaluations can be foundhere. In addition to Nemo Skills, the evaluations also used dedicated open-source packaged containers for Tau-2 Bench (default prompt), Terminal Bench Hard (48 tasks), ScaleAI Multi Challenge Multi-turn Instruction Following, and Ruler.

The following benchmarks are not onboarded yet in our open source tools and for these we used either their official open source implementation or otherwise an internal scaffolding that we plan to open source in the future: SWE Bench Verified (OpenHands).

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#deployment-geographyDeployment Geography:

Global

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#use-case-Use Case:

NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 is a general purpose reasoning and chat model intended to be used in English, Code, and supported multilingual contexts. This model is optimized for collaborative agents and high-volume workloads. It is intended to be used by developers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. This model is also suitable for complex instruction-following tasks and long-context reasoning.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#release-date–Release Date:

July 6, 2026 viaHugging Face

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#referencessReferences(s):

• [2411.19146] Puzzle: Distillation-Based NAS for Inference-Optimized LLMs
• [2604.12374] Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#model-architectureModel Architecture:

• **Architecture Type:**Mamba2-Transformer Hybrid Latent Mixture of Experts (LatentMoE) with Multi-Token Prediction (MTP)
• **Network Architecture:**ModifiedNemotron-3-Super-120B-A12B-NVFP4architecture with smaller Mamba SSM state size, varying number of active experts per layer and varying expert intermediate channel size across layers.
• **Number of model parameters:**75B Total / 9.3B Active

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#model-designModel Design

Puzzle-75B-A9B is a compressed variant of Nemotron-3-Super optimized for interactive deployment. We designed the model to maximize server throughput under high user throughput constraints.

The model was constructed using a multi-stage pipeline that combines the Iterative Puzzle compression framework with knowledge distillation, reinforcement learning, quantization, and Multi-Token Prediction head. The compression process jointly optimizes heterogeneous MoE pruning, active parameter budget, and Mamba pruning to improve inference efficiency while preserving model quality. Attention layers are left unchanged because the parent model is already KV-cache efficient.

Compression is applied to three architectural dimensions:

• Heterogeneous MoE Channel Pruning: Routed expert intermediate dimensions are pruned non-uniformly across MoE layers. The parent routed expert intermediate size of 2688 is reduced to a layer-dependent range of 1280-2688, preserving more capacity in sensitive layers while pruning more aggressively elsewhere.
• Heterogeneous Active Expert Reduction: The number of activated routed experts per token is reduced from 22 in the parent model to a layer-dependent range of 4-18. This reduces active parameters and improves efficiency in compute-bound inference regimes such as prefill and large-batch decoding.
• Mamba SSM State Pruning: The Mamba SSM state size is reduced from 128 to 96 channels. This reduces Mamba cache I/O and improves decode-stage efficiency, especially at larger batch sizes.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#training-and-optimization-procedureTraining and Optimization Procedure

Puzzle-75B-A9B is produced through a post-training compression and recovery pipeline starting from Nemotron-3-Super. The pipeline combines Iterative Puzzle compression, knowledge distillation, reinforcement learning recovery, post-training quantization, and continued MTP training.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#stage-1-iterative-puzzle-compressionStage 1: Iterative Puzzle Compression

The model is constructed through three compression-and-recovery stages. Each stage prunes the model to a certain intermediate target budget and then performs a short knowledge distillation recovery phase before the next compression step.

In the first stage, MoE weights are reduced to 75% of the teacher capacity, and the Mamba SSM state size is reduced to 75% of the teacher size. The resulting model is recovered with 24B tokens of knowledge distillation. In the second stage, MoE weights are further reduced to 60% of the teacher capacity, followed by 43.2B tokens of knowledge distillation recovery. In the final stage, the activated routed-expert budget (MoE top-k) is constrained to 50% of the teacher budget, with Puzzle allocating this budget heterogeneously across layers. The resulting model is recovered with 52.8B tokens of knowledge distillation.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#stage-2-long-context-knowledge-distillation-recoveryStage 2: Long-Context Knowledge Distillation Recovery

After architecture selection, the compressed model undergoes additional knowledge distillation from Nemotron-3-Super to recover quality lost during compression and recover long-context capability.

Training uses a mixture of 30% pretraining data and 70% supervised fine-tuning data. During the Iterative Puzzle stages, knowledge distillation is performed at 32Ki sequence length. The final recovery phase extends distillation to longer contexts, first at 128Ki and then at 512Ki sequence length, using up to 100B training tokens per phase and a global batch size of 16Mi tokens.

Software used for knowledge distillation:Megatron-BridgeandMegatron-LM.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#stage-3-reinforcement-learning-rl-recoveryStage 3: Reinforcement Learning (RL) Recovery

Following knowledge distillation, the model undergoes reinforcement learning recovery focused primarily on software-engineering and agentic capabilities, which are especially sensitive to compression.

The RL stage follows the Nemotron-3-Super software-engineering RL pipeline (SWE-RL). It includes single-step tool-use comparison training and end-to-end sandbox RL, where agents interact with isolated execution environments over multiple turns. Multiple RL runs are trained with different learning rates, and the final checkpoint is obtained through weight averaging across selected runs.

Software used for reinforcement learning:NeMo-RL

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#stage-4-deployment-optimizationStage 4: Deployment Optimization

The resulting checkpoint is further prepared for deployment using post-training quantization. FP8 checkpoints target Hopper-class GPUs, while NVFP4 checkpoints target Blackwell-class GPUs. The model also uses continued MTP training to improve speculative decoding acceptance length and increase serving throughput.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#inputInput

• **Input Type(s):**Text
• **Input Format(s):**String
• **Input Parameters:**One-Dimensional (1D): Sequences
• **Other Properties Related to Input:**Maximum context length up to 1M tokens. Supported languages include: English, French, German, Italian, Japanese, Spanish, and Chinese

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#outputOutput

• **Output Type(s):**Text
• **Output Format:**String
• **Output Parameters:**One-Dimensional (1D): Sequences
• **Other Properties Related to Output:**Maximum context length up to 1M tokens

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.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#software-integrationSoftware Integration:

• Runtime Engine(s): Hugging Face Transformers, vLLM
• Supported Hardware Microarchitecture Compatibility: NVIDIA Blackwell, NVIDIA Hopper
• 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.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#model-versionsModel Version(s)

• v1.0 - GA

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#quick-start-guideQuick Start Guide

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#servingServing

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#vllmvLLM

To deploy the Nemotron Labs 3 PuzzleNVFP4checkpoint on NVIDIABlackwellGPUs, use the following command:

• With MTP:vllm serve "$path" \ --served-model-name "$model" \ --port "$port" \ --tensor-parallel-size "$tp" \ --enable-expert-parallel \ --async-scheduling \ --trust-remote-code \ --mamba-backend flashinfer \ --mamba_ssm_cache_dtype float16 \ --enable-mamba-cache-stochastic-rounding \ --mamba-cache-philox-rounds 5 \ --speculative-config "{\"method\":\"mtp\",\"num_speculative_tokens\":${num_speculative_tokens}}" \ --tool-call-parser qwen3_coder \ --reasoning-parser nemotron_v3 \ --enable-auto-tool-choice
• Without MTP:vllm serve "$path" \ --served-model-name "$model" \ --port "$port" \ --tensor-parallel-size "$tp" \ --enable-expert-parallel \ --mamba_ssm_cache_dtype float16 \ --enable-mamba-cache-stochastic-rounding \ --mamba-cache-philox-rounds 5 \ --async-scheduling \ --trust-remote-code \ --mamba-backend flashinfer \ --tool-call-parser qwen3_coder \ --reasoning-parser nemotron_v3 \ --enable-auto-tool-choice

Notes:

• Tested onvLLM v0.20.0.
• NVIDIA recommends settingtpto2or4.
• For MTP,num\_speculative\_tokens=3is the recommended default (best throughput at typical BS);5or7may be beneficial for low-batch / latency-sensitive deployments.
• For very long generation scenarios, it is reccomeneded to use\-\-api\-server\-count 4.\-\-no\-enable\-chunked\-prefillcan be used to increase throughput, but potentially reduce reponsiveness.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#api-clientAPI Client

The examples below use the OpenAI-compatible client.

NOTE: For coding agents add the following to the API call -extra\_body=\{“chat\_template\_kwargs”: \{“force\_nonempty\_content”: True\}

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
MODEL = "nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4"

Reasoning ON (default)

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "Write a haiku about GPUs"}],
    max_tokens=16000,
    temperature=1.0,
    top_p=0.95,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}}
)
print(response.choices[0].message.content)

Reasoning OFF

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "What is the capital of Japan?"}],
    max_tokens=16000,
    temperature=1.0,
    top_p=0.95,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}}
)
print(response.choices[0].message.content)

Low-effort reasoning

Uses significantly fewer reasoning tokens than full thinking mode. Recommended as a starting point before tuning explicit token budgets.

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "What is the capital of Japan?"}],
    max_tokens=16000,
    temperature=1.0,
    top_p=0.95,
    extra_body={"chat_template_kwargs": {"enable_thinking": True, "low_effort": True}}
)
print(response.choices[0].message.content)

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#transformersTransformers

We recommend using Transformers ≥ 5.3.0.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4")
model = AutoModelForCausalLM.from_pretrained(
    "nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

Please note that the model supports up to a 1M context size, although the default context size in the Hugging Face configuration is 256k due to higher VRAM requirements.

Here is an example of generating outputs with reasoning enabled (the default):

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)

if not isinstance(tokenized_chat, torch.Tensor):
    input_ids = tokenized_chat["input_ids"]
else:
    input_ids = tokenized_chat

outputs = model.generate(
    input_ids,
    max_new_tokens=50,
    temperature=1.0,
    top_p=0.95,
    eos_token_id=tokenizer.eos_token_id
)

print(tokenizer.decode(outputs[0]))

To disable reasoning, addenable\_thinking=Falsetoapply\_chat\_template\(\). By default,enable\_thinkingis set toTrue.

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

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#training-and-evaluation-datasetsTraining and Evaluation Datasets

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#trainingTraining

**Data Modality:**Text **The total size:**15,573,172,908,990 Tokens **Total number of datasets:**153 **Dataset partition:**Training [100%], testing [0%], validation [0%] **Time period for training data collection:**2013 to February 24, 2026 **Time period for testing data collection:**2013 to February 24, 2026 **Time period for validation data collection:**2013 to February 24, 2026 **Data Collection Method by dataset:**Hybrid: Automated, Human, Synthetic **Labeling Method by dataset:**Hybrid: Automated, Human, Synthetic

NVIDIA-Nemotron-3-Super-120B-A12B 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-Super-120B-A12B of high-quality curated and synthetically-generated data. Primary languages used for post-training include English, French, German, Italian, Japanese, Spanish, and Chinese.

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 Super*.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#additional-training-data-for-puzzle-75b-a9bAdditional Training Data for Puzzle-75B-A9B

NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 is initialized from NVIDIA-Nemotron-3-Super-120B-A12B and therefore inherits the parent model’s pre-training and post-training data exposure described above.

For compression recovery, the model is trained with knowledge distillation on a mixed dataset consisting of 30% pretraining data and 70% supervised fine-tuning data from theNemotron-3-Nanotraining pipeline. Distillation uses NVIDIA-Nemotron-3-Super-120B-A12B-BF16 as the teacher model and is performed during both the Iterative Puzzle compression stages and the subsequent long-context recovery stages.

The long-context recovery data is used at 128Ki and 512Ki sequence lengths to recover long-context capabilities after compression.

After knowledge distillation, the model undergoes reinforcement learning recovery using software-engineering and agentic task data from theNemotron-3-SuperRL pipeline, including single-step tool-use comparison data and end-to-end sandbox RL environments.

Click to explore the full dataset catalogue used for training#### https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#base-pre-training-corpus-nemotron-3-foundationBase Pre-Training Corpus (Nemotron 3 Foundation)

The foundation of the model is trained on theNemotron-3-Nanocorpus, comprising the following collections:

Dataset CollectionToken CountsDescriptionNemotron-CC-v2&v2.19.13TA massive collection of English web data filtered from Common Crawl, including 2.5T+ tokens of new organic, translated, and synthetically rephrased content.Nemotron-CC-Code-v1427.9BHigh-quality code tokens extracted from Common Crawl using the Lynx + LLM pipeline to preserve structure and equations.Nemotron-Pretraining-Code-v1&v21.09TCurated GitHub code references with multi-stage filtering, deduplication, and large-scale synthetic code data.Nemotron-CC-Math-v1133.3BHigh-quality math pre-training dataset preserving LaTeX formatting and mathematical structures.Nemotron-Pretraining-Specialized-v1336.4BSynthetic datasets targeting specialized domains such as STEM reasoning and scientific coding.

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#public-datasetsPublic Datasets

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#crawled-and-scraped-from-online-sources-by-nvidiaCrawled 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 thetechnical report).

DatasetModalityDataset SizeCollection PeriodCollecting OrganisationEnglish Common CrawlText3.36T4/8/2025NVIDIA Advanced Deep Learning ResearchEnglish Common Crawl 1.1TextNot disclosed10/2/2025NVIDIA Advanced Deep Learning ResearchMultilingual Common CrawlText812.7B5/1/2025NVIDIA Advanced Deep Learning ResearchGitHub CrawlText747.4B4/29/2025NVIDIA Advanced Deep Learning Research

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#private-non-publicly-accessible-datasets-of-third-partiesPrivate Non-publicly Accessible Datasets of Third Parties

DatasetModel(s) usedGlobal RegulationUnknownTAUS Translation MemoryUnknownScale HLEUnknownHackerRank CodingUnknownRL data for SearchGemini 3; GPT-5 *

• Models used for prompt generation only

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#private-non-publicly-accessible-datasets-by-nvidiaPrivate Non-publicly Accessible Datasets by NVIDIA

DatasetModel(s) usedSimple Minesweeper-Simple Sudoku-Multitool Typewriter Hard-Machine Translation of News Commentary and TAUS Translation Memory-Machine Translation of STEM -Qwen2.5-14B-InstructCompetitive Coding RL data from Nemotron Cascade-Long context RL-Single-step SWE RL for patch generation-OpenHands SWE-

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#nvidia-sourced-synthetic-datasetsNVIDIA-Sourced Synthetic Datasets

DatasetModalityDataset SizeSeed DatasetModel(s) used for generationNemotron-Pretraining-Formal-LogicText128,022,285Nemotron PersonasQwen3-235B-A22B-Thinking-2507Nemotron-Pretraining-EconomicsText73,374,154-Qwen3-235B-A22B-Thinking-2507Nemotron-Pretraining-Multiple-ChoiceText1,609,214,470MMLU Auxiliary TrainDeepSeek-V3;Qwen3-235B-A22BNemotron-Pretraining-Code-ConceptsText7,294,510,156-gpt-oss-20b;gpt-oss-120bNemotron-Pretraining-Unconditional-AlgorithmicText196,492,899-gpt-oss-120b;Qwen3-235B-A22BSynthetic Tasks from DeepSeek-V3 and Qwen3-235B-A22BText6.7Btrain splits of Into the Unknown; AI2 ARC (AI2 Reasoning Challenge); BLiMP (Benchmark of Linguistic Minimal Pairs); CommonSenseQA; GLUE; HeadQA; Hendrycks Ethics; Memo Trap; modus-tollens; NeQA; pattern-matching-suppression; mastermind_24_mcq_random; mastermind_24_mcq_close; quote-repetition; redefine-math; Repetitive Algebra; sig-figs; MMLU-Pro; MC-TACO; MedConceptsQA; MMLU_dataset; OpenbooksQA; PIQA (Physical Interaction Question Answering); SocialIQA; SuperGLUE; tinyAI2_arc; tinyMMLU; tinyWinogrande; TruthfulQA; WebQuestions; Winogrande; GPQA; MBPPDeepSeek v3;Qwen3-235B-A22BSynthetic Art of Problem Solving from DeepSeek-R1Text40BArt of Problem Solving;American Mathematics Competitions 8;American Mathematics Competitions 10;DeepSeek-R1Synthetic Moral Stories and Social Chemistry from Mixtral-8x22B-v0.1Text327Msocial-chemestry-101;Moral StoriesMixtral-8x22B-v0.1Synthetic Social Sciences seeded with OpenStax from DeepSeek-V3, Mixtral-8x22B-v0.1, and Qwen2.5-72BText83.6MOpenStax - CC BY-SA subsetDeepSeek-V3;Mixtral-8x22B-v0.1;Qwen2.5-72BSynthetic Health Sciences seeded with OpenStax from DeepSeek-V3, Mixtral-8x22B-v0.1, and Qwen2.5-72BText9.7MOpenStax - CC BY-SA subsetDeepSeek-V3;Mixtral-8x22B-v0.1;Qwen2.5-72BSynthetic STEM seeded with OpenStax, Open Textbook Library, and GSM8K from DeepSeek-R1, DeepSeek-V3, DeepSeek-V3-0324, and Qwen2.5-72BText175MOpenStax - CC BY-SA subset;GSM8K;Open Textbook Library - CC BY-SA & GNU subsetDeepSeek-R1,DeepSeek-V3;DeepSeek-V3-0324;Qwen2.5-72BNemotron-PrismMathText4.6BBig-Math-RL-Verified;OpenR1-Math-220kQwen2.5-0.5B-instruct,Qwen2.5-72B-Instruct;DeepSeek-R1-Distill-Qwen-32BSynthetic Question Answering Data from Papers and Permissible Books from Qwen2.5-72B-InstructText350MarXiv;National Institutes of Health ExPorter;BioRxiv;PMC Article;USPTO Backgrounds;peS2o; Global Regulation;CORE;PG-19;DOAB CC BY & CC BY-SA subset;NDLTDQwen2.5-72B-InstructRefreshedNemotron-MINDfrom phi-4Text73BCommon Crawlphi-4Nemotron-CC-Math-4plusText52.3BCommon Crawlphi-4Nemotron-CC-Math-3Text80.9BCommon Crawlphi-4Synthetic AGIEval seeded with AQUA-RAT, LogiQA, and AR-LSAT from DeepSeek-V3 and DeepSeek-V3-0324Text4.0BAQUA-RAT;LogiQA;AR-LSATDeepSeek-V3;DeepSeek-V3-0324Synthetic AGIEval seeded with AQUA-RAT, LogiQA, and AR-LSAT from Qwen3-30B-A3BText4.2BAQUA-RAT;LogiQA;AR-LSATQwen3-30B-A3BSynthetic Art of Problem Solving from Qwen2.5-32B-Instruct, Qwen2.5-Math-72B, Qwen2.5-Math-7B, and Qwen2.5-72B-InstructTextArt of Problem Solving;American Mathematics Competitions 8;American Mathematics Competitions 10;GSM8K;PRM800KQwen2.5-32B-Instruct;Qwen2.5-Math-72B;Qwen2.5-Math-7B;Qwen2.5-72B-InstructSynthetic MMLU Auxiliary Train from DeepSeek-R1Text0.5BMMLU Auxiliary TrainDeepSeek-R1Synthetic Long Context Continued Post-Training Data from Papers and Permissible Books from Qwen2.5-72B-InstructTextarXiv;National Institutes of Health ExPorter;BioRxiv;PMC Article;USPTO Backgrounds;peS2o; Global Regulation;CORE;PG-19;DOAB CC BY & CC BY-SA subset;NDLTDQwen2.5-72B-InstructSynthetic Common Crawl from Qwen3-30B-A3B and Mistral-Nemo-12B-InstructText415.8BCommon CrawlQwen3-30B-A3B;Mistral-NeMo-12B-InstructSynthetic Multilingual Data from Common Crawl from Qwen3-30B-A3BTextCommon CrawlQwen3-30B-A3BSynthetic Multilingual Data from Wikimedia from Qwen3-30B-A3BTextWikimediaQwen3-30B-A3BSynthetic Math Data from Wikimedia from Nemotron-4-340B-InstructText-Nemotron-4-340B-InstructSynthetic Common Crawl Code from phi-4Text427.9BCommon Crawlphi-4Synthetic Scientific Coding from Qwen3-235B-A22BText1.2BWikimediaQwen3-235B-A22BTool Calling DataText26.2BQwen3-235B-A22B-2507;gpt-oss-120bSynthetic Essential-Web from QwQ-32BText28.1BEssential-WebQwQ-32BTranslated Synthetic CrawlText389.9BCommon CrawlQwen3-30B-A3BTranslated Synthetic WikipediaText7.9BWikimediaQwen3-30B-A3BSynthetic Art of Problem Solving from gpt-oss-120b and Qwen2.5-32B-InstructTextUndisclosedArt of Problem Solving;American Mathematics Competitions 8;American Mathematics Competitions 10gpt-oss-120b;Qwen2.5-32B-InstructSynthetic Stack Exchange from gpt-oss-120b and Qwen2.5-32B-InstructTextUndisclosedStack Exchangegpt-oss-120b;Qwen2.5-32B-InstructSynthetic OpenCodeReasoning from DeepSeek-R1-0528TextUndisclosedOpenCodeReasoningDeepSeek-R1-0528Synthetic HackerRank Coding from DeepSeek-R1-0528TextUndisclosedHackerRank Coding DatasetDeepSeek-R1-0528Synthetic SWE-Gym from Qwen3-Coder-480B-A35B-InstructTextUndisclosedSWE-GymQwen3-Coder-480B-A35B-InstructSynthetic Art of Problem Solving and Stack Exchange from gpt-oss-120b, Qwen2.5-32B-Instruct, and Goedel-Prover-V2-32BTextUndisclosedArt of Problem Solving;American Mathematics Competitions 8;American Mathematics Competitions 10;Stack Exchangegpt-oss-120b;Qwen2.5-32B-Instruct;Goedel-Prover-V2-32BSynthetic Multilingual Science and Code data from DeepSeek-R1, DeepSeek-R1-0528, Qwen2.5-32B-Instruct, and Qwen3-235B-A22B, translated with Qwen2.5-32B-Instruct and Qwen2.5-14B-InstructTextUndisclosedStack Exchange;SCP-116K;LIMO;TACO; Code Contest; CodeforcesDeepSeek-R1;DeepSeek-R1-0528;Qwen2.5-32B-Instruct;Qwen3-235B-A22B;Synthetic Safety from DeepSeek-R1-0528, gpt-oss-120b and Mixtral-8x7B-v0.1TextUndisclosedNemotron Content Safety Dataset V2;Gretel Synthetic Safety Alignment Dataset;RedTeam-2K;Malicious Tasks;Nemotron-Personas-USADeepSeek-R1-0528;gpt-oss-120b;Mixtral-8x7B-v0.1Synthetic STEM from Qwen3-235B-A22B-Instruct-2507 and gpt-oss-120bTextUndisclosedarXiv;National Institutes of Health ExPorter;BioRxiv;PMC Article;USPTO Backgrounds;peS2o; Global Regulation;CORE;PG-19;DOAB CC BY & CC BY-SA subset;NDLTDQwen3-235B-A22B-Instruct-2507;gpt-oss-120bSynthetic KernelBook from DeepSeek-R1-0528TextUndisclosedKernelBookDeepSeek-R1-0528Synthetic Tool Calling from Qwen3-235B-A22B-Thinking-2507 and Qwen3-Next-80B-A3B-ThinkingTextUndisclosedToolBench;glaive-function-calling-v2;APIGen Function-Calling;Nemotron-Personas-USAQwen3-235B-A22B-Thinking-2507;Qwen3-Next-80B-A3B-ThinkingSynthetic Chat from gpt-oss-120b, Mixtral-8x22B-Instruct-v0.1, Qwen3-235B-A22B-Instruct-2507 , and Qwen3-235B-A22B-Thinking-2507TextUndisclosedC4;LMSYS-Chat-1M;ShareGPT;GSM8K;PRM800K;FinQA;WikiTableQuestions;Riddles;glaive-function-calling-v2;SciBench;tigerbot-kaggle-leetcodesolutions-en-2k;OpenBookQA;Advanced Reasoning Benchmark; Software Heritage;Khan Academy Math Keywords;WildChat-1M;Nemotron-Personas-USAgpt-oss-120b;Mixtral-8x22B-Instruct-v0.1;Qwen3-235B-A22B-Instruct-2507;Qwen3-235B-A22B-Thinking-2507Synthetic Long Context from Qwen3-235B-A22B-Instruct-2507TextUndisclosedCORE;PG-19;DOAB CC BY & CC BY-SA subset;NDLTDQwen3-235B-A22B-Instruct-2507Synthetic Tool Use Interactive Agent from gpt-oss-120b, DeepSeek-R1-0528, Qwen3-32B, and Qwen3-235B-A22B-Thinking-2507TextUndisclosedNVIDIA Internalgpt-oss-120b;DeepSeek-R1-0528;Qwen3-32B; andQwen3-235B-A22B-Thinking-2507Synthetic STEM from Qwen3-235B-A22B-Thinking-2507TextUndisclosedICHO-IPH0;Physics Big; Scale HLE;OpenMathReasoning;OpenCodeReasoningQwen3-235B-A22B-Thinking-2507Synthetic DocFinQA and SWE-smith from Qwen3-Coder-480B-A35B-Instruct and Kimi-K2-ThinkingTextUndisclosedDocFinQA;SWE-smithQwen3-Coder-480B-A35B-Instruct;Kimi-K2-ThinkingSynthetic Math from gpt-oss-120b and Qwen2.5-32B-InstructTextUndisclosed-gpt-oss-120b;Qwen2.5-32B-InstructSynthetic Essential-Web from gpt-oss-120bTextUndisclosedEssential-Webgpt-oss-120bSynthetic Scale HLE from gpt-oss-120bTextUndisclosedScale HLEgpt-oss-120bSynthetic CDQuestions from gpt-oss-120bTextUndisclosedCDQuestionsgpt-oss-120bSynthetic Stack Exchange from gpt-oss-120bTextUndisclosedStack Exchangegpt-oss-120bSynthetic GPQA from gpt-oss-120b and Qwen2.5-32B-InstructTextUndisclosedStack Exchangegpt-oss-120b;Qwen2.5-32B-InstructSynthetic Vedantu from gpt-oss-120bTextUndisclosedVedantugpt-oss-120bSynthetic SWE-Gym and R2E-Gym-Subset from Qwen3-Coder-480B-A35B-InstructTextUndisclosedSWE-Gym;R2E-Gym-SubsetQwen3-Coder-480B-A35B-InstructSynthetic SWE-Gym from Qwen3-Coder-480B-A35B-InstructTextUndisclosedSWE-GymQwen3-Coder-480B-A35B-InstructSynthetic SWE-Gym and R2E-Gym-Subset from DeepSeek-R1-0528TextUndisclosedSWE-Gym;R2E-Gym-SubsetDeepSeek-R1-0528Synthetic HelpSteer, LMSYS-Chat-1M, and Nemotron-Personas-USA from gpt-oss-120b, Qwen3-235B-A22B-Instruct-2507, and Qwen3-235B-A22B-Thinking-2507TextUndisclosedHelpSteer2;HelpSteer3;LMSYS-Chat-1M;Nemotron-Personas-USAgpt-oss-120b;Qwen3-235B-A22B-Instruct-2507;Qwen3-235B-A22B-Thinking-2507Synthetic Structured Outputs from Qwen3-30B-A3B-Instruct-2507, Qwen3-30B-A3B-Thinking-2507, Qwen3-235B-A22B-Instruct-2507, and Qwen3-235B-A22B-Thinking-2507TextUndisclosed-Qwen3-30B-A3B-Instruct-2507;Qwen3-30B-A3B-Thinking-2507;Qwen3-235B-A22B-Instruct-2507;Qwen3-235B-A22B-Thinking-2507Synthetic Search STEM MCQ from Qwen3-235B-A22B and DeepSeek-R1-0528TextUndisclosed-Qwen3-235B-A22B;DeepSeek-R1-0528Synthetic Search STEM OPENQ from DeepSeek-R1-0528TextUndisclosed-DeepSeek-R1-0528Synthetic OpenSTEM from Qwen2.5-32B-Instruct and DeepSeek-R1-0528TextUndisclosed-Qwen2.5-32B-Instruct;DeepSeek-R1-0528Synthetic MCQ from Qwen2.5-32B-Instruct and DeepSeek-R1-0528TextUndisclosed-Qwen2.5-32B-Instruct;DeepSeek-R1-0528Synthetic MCQ10 from DeepSeek-R1-0528TextUndisclosed-DeepSeek-R1-0528Synthetic MCQ4 from Qwen3-235B-A22B, DeepSeek-R1-0528, and Qwen3-235B-A22B-Instruct-2507TextUndisclosed-Qwen3-235B-A22B;DeepSeek-R1-0528;Qwen3-235B-A22B-Instruct-2507Synthetic OpenMathReasoning from gpt-oss-120b and Qwen2.5-32B-InstructTextUndisclosedOpenMathReasoninggpt-oss-120b;Qwen2.5-32B-InstructSynthetic Offline Search MCQA HLE from DeepSeek-R1-0528TextUndisclosed-DeepSeek-R1-0528Synthetic Offline Search MCQA GPQA from Qwen3-235B-A22B and DeepSeek-R1-0528TextUndisclosed-Qwen3-235B-A22B;DeepSeek-R1-0528Synthetic Human Preference from QwQ-32B, Qwen3-30B-A3B, Qwen3-235B-A22B, Qwen3-235B-A22B-Instruct-2507, Mistral-Small-3.1-24B-Instruct-2503, Mistral-Small-3.2-24B-Instruct-2506, MiniMax-M1-80k, MiniMax-M1-40k, Kimi-K2-Instruct, DeepSeek-V3-0324, DeepSeek-R1-0528TextUndisclosed-QwQ-32B;Qwen3-30B-A3B;Qwen3-235B-A22B;Qwen3-235B-A22B-Instruct-2507;Mistral-Small-3.1-24B-Instruct-2503;Mistral-Small-3.2-24B-Instruct-2506;MiniMax-M1-80k;MiniMax-M1-40k;Kimi-K2-Instruct;DeepSeek-V3-0324;DeepSeek-R1-0528Synthetic WildChat-1M and arena-human-preference-140k from DeepSeek-R1, gemma-2-2b-it, gemma-3-27b-it, gpt-oss-20b, gpt-oss-120b, Mistral-7B-Instruct-v0.3, Mixtral-8x22B-Instruct-v0.1, Nemotron-4-340B-Instruct, NVIDIA-Nemotron-Nano-9B-v2, Phi-4-mini-instruct, Phi-3-small-8k-instruct, Phi-3-medium-4k-instruct, Qwen3-235B-A22B, QwQ-32BTextUndisclosedWildChat-1M;arena-human-preference-140kDeepSeek-R1;gemma-2-2b-it;gemma-3-27b-it;gpt-oss-20b;gpt-oss-120b;Mistral-7B-Instruct-v0.3;Mixtral-8x22B-Instruct-v0.1;Nemotron-4-340B-Instruct;NVIDIA-Nemotron-Nano-9B-v2;Phi-4-mini-instruct;Phi-3-small-8k-instruct;Phi-3-medium-4k-instruct;Qwen3-235B-A22B;QwQ-32BSynthetic Safety from DeepSeek-R1-0528, gpt-oss-120b, DeepSeek-R1-Distill-Qwen-7B, and Mixtral-8x7B-v0.1TextUndisclosedNemotron Content Safety Dataset V2;Gretel Synthetic Safety Alignment Dataset;RedTeam-2K;Malicious Tasks;DeepSeek-R1-0528;gpt-oss-120b;DeepSeek-R1-Distill-Qwen-7B;Qwen3-30B-A3B-Thinking-2507;Qwen3-235B-A22B-Instruct-2507;Mixtral-8x7B-v0.1Synthetic Code from Qwen3-32BTextUndisclosedEnglish Common Crawl; English Common Crawl 1.1Qwen3-32BSynthetic OpenCodeReasoning from DeepSeek-R1TextUndisclosedOpenCodeReasoningDeepSeek-R1Synthetic LIMO from DeepSeek-R1-0528TextUndisclosedLIMODeepSeek-R1-0528Synthetic SCP from DeepSeek-R1-0528TextUndisclosedSCP-116KDeepSeek-R1-0528Synthetic Stack Exchange from DeepSeek-R1-0528TextUndisclosedStack ExchangeDeepSeek-R1-0528Synthetic Common Crawl from Qwen3-30B-A3BTextUndisclosedCommon CrawlQwen3-30B-A3BSynthetic Wikipedia from Qwen3-30B-A3BTextUndisclosedWikimediaQwen3-30B-A3BSynthetic Essential-Web from Qwen3-30B-A3B and Qwen3-235B-A22B-Thinking-2507TextUndisclosedEssential-WebQwen3-30B-A3B;Qwen3-235B-A22B-Thinking-2507Synthetic Textbook Math from Qwen3-30B-A3B, Qwen3-235B-A22B, phi-4TextUndisclosedCommon Crawl;FineMathQwen3-30B-A3B;Qwen3-235B-A22B;phi-4Synthetic Math and Code from DeepSeek-R1 and DeepSeek-R1-0528TextUndisclosedMagicoder-Evol-Instruct-110K;opc-sft-stage2;TACO;OpenCodeReasoning;OpenMathReasoning;NuminaMath CoTDeepSeek-R1;DeepSeek-R1-0528Synthetic Nemotron-Personas-USA from gpt-oss-120b and Qwen3-8BTextUndisclosedNemotron-Personas-USAgpt-oss-120b;Qwen3-8BSynthetic Text-To-SQLTextUndisclosed-gpt-oss-120bSynthetic Agentless SWETextUndisclosedSWE-Bench-Train;SWE-Fixer-Train;SWE-reBench;SWE-smithDeepSeek-R1-0528Synthetic Search Graph WalkTextUndisclosed-MiniMax-M2Synthetic CUDA 100kTextUndisclosedKernelBook;HuggingFace Transformers;FlashInferDeepSeek-R1-0528;gpt-oss-120bSynthetic SafetyTextUndisclosedNemotron Content Safety Dataset V2;Gretel Synthetic Safety Alignment Dataset;RedTeam-2K;HarmfulTasksgpt-oss-120b;NVIDIA-Nemotron-Nano-9B-v2;gemma-3-4b-itSynthetic Agentic Diverse DomainsTextUndisclosed-DeepSeek-R1-0528;Qwen3-235B-A22B-Thinking-2507;Qwen3-235B-A22B-Instruct-2507;Qwen3-32B;gpt-oss-120b;DeepSeek-V3.2Synthetic SWE UnverifiedTextUndisclosed-gpt-oss-120b;Qwen3-Coder-480B-A35B-Instruct;GLM-4.7-FlashSynthetic Scale HLE from Deepseek-V3TextUndisclosedScale HLEDeepSeek-V3-0324Synthetic CDQuestions from Deepseek-V3TextUndisclosedCDQuestionsDeepSeek-V3-0324Synthetic Stack Exchange from Deepseek-V3TextUndisclosedStack ExchangeDeepSeek-V3-0324Synthetic GPQA from Deepseek-V3TextUndisclosedStack ExchangeDeepSeek-V3-0324Synthetic Vedantu from Deepseek-V3TextUndisclosedVedantuDeepSeek-V3-0324Synthetic Tool Call Schema for RLTextUndisclosedToolBench;glaive-function-calling-v2;APIGen Function-Calling;Nemotron-Personas-USAQwen3-235B-A22B-Thinking-2507;Qwen3-Next-80B-A3B-ThinkingSynthetic Data for SearchTextUndisclosedWikimediaMiniMax-M2Synthetic Instruction Following for RLTextUndisclosed-NVIDIA-Nemotron-Nano-9B-v2;Qwen3-235B-A22B-Thinking-2507Synthetic Conversational Agentic Tool-Use RLTextUndisclosed-DeepSeek-V3.2;DeepSeek-R1-0528;Qwen3-235B-A22B-Thinking-2507;Qwen3-32B;gpt-oss-120b;Qwen3-235B-A22B-Instruct-2507Synthetic Terminal Pivot RLTextUndisclosedSWE-smith;Nemotron-Cascade-RL-SWE; Vendor suppliedDeepSeek-V3.2;Qwen3-Coder-480B-A35B-Instruct;Kimi-K2.5;Qwen3-235B-A22B-Instruct-2507

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#evaluation-datasetEvaluation Dataset

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

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#inferenceInference:

**Acceleration Engine:**vLLM

Test Hardware:

• 1× NVIDIA H100-80GB
• 8× NVIDIA H100-80GB
• 8× NVIDIA B200

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4#ethical-considerationsEthical 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.

We advise against circumvention of any provided safety guardrails contained in the Model without a substantially similar guardrail appropriate for your use case. For more details:SafetyandExplainabilitySubcards.

For more detailed information on ethical considerations for this model, please see the Model Card++Bias, andPrivacySubcards.

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