NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4
NVIDIA's 75B-parameter MoE model with 9B active parameters, NVFP4-quantized for text generation.
Base model
Model Description
Description:
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.
License/Terms of Use
Governing Download Terms: Use of this model is governed by the OpenMDW License Agreement, version 1.1 (OpenMDW-1.1).
This project is currently not accepting contributions.
Benchmarks
| Benchmark | Nemotron-Labs-3-Puzzle-75B-A9B-BF16 | Nemotron-Labs-3-Puzzle-75B-A9B-FP8 | Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 |
|---|---|---|---|
| General Knowledge | |||
| MMLU-Pro | 82.4 | 82.0 | 82.2 |
| Reasoning | |||
| AIME25 (no tools) | 89.7 | 89.4 | 89.9 |
| HMMT Feb25 (no tools) | 93.4 | 92.7 | 92.9 |
| HMMT Feb25 (with tools) | 93.9 | 93.6 | 93.1 |
| GPQA (no tools) | 78.6 | 77.8 | 78.0 |
| GPQA (with tools) | 79.5 | 80.6 | 78.2 |
| LiveCodeBench (v5 2024-07↔2024-12) | 81.1 | 80.5 | 79.9 |
| SciCode (subtask) | 40.6 | 39.6 | 40.3 |
| HLE (no tools) | 16.5 | 16.0 | 15.7 |
| Agentic | |||
| Terminal Bench (hard subset) | 24.0 | 22.9 | 23.4 |
| TauBench V2 | |||
| Airline | 55.8 | 54.5 | 55.7 |
| Retail | 63.2 | 63.4 | 63.7 |
| Telecom | 61.5 | 61.3 | 60.3 |
| Average | 60.2 | 59.7 | 59.9 |
| Chat & Instruction Following | |||
| IFBench (prompt) | 71.9 | 71.9 | 71.3 |
| Scale AI Multi-Challenge | 56.6 | 55.4 | 55.9 |
| Arena-Hard-V2 | 68.6 | 69.8 | 69.0 |
| Long Context | |||
| AA-LCR | 56.9 | 56.6 | 57.1 |
| RULER @ 256k | 95.1 | 95.3 | 95.3 |
| RULER @ 512k | 94.2 | 94.5 | 94.8 |
| RULER @ 1M | 92.2 | 92.4 | 93.2 |
| Multilingual | |||
| MMLU-ProX (avg over langs) | 77.5 | 77.1 | 76.5 |
| WMT24++ (en→xx) | 85.2 | 85.2 | 85.1 |
All evaluation results were collected via Nemo Evaluator SDK and for most benchmarks, the Nemo Skills Harness. For reproducibility purposes, more details on the evaluation settings can be found in the Nemo Evaluator SDK configs folder and the reproducibility tutorial for Nemotron 3 Super. The open source container on Nemo Skills packaged via NVIDIA's Nemo Evaluator SDK used for evaluations can be found here. 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).
Deployment Geography:
Global
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.
Release Date:
July 6, 2026 via Hugging Face
References(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
Model Architecture:
- Architecture Type: Mamba2-Transformer Hybrid Latent Mixture of Experts (LatentMoE) with Multi-Token Prediction (MTP)
- Network Architecture: Modified Nemotron-3-Super-120B-A12B-NVFP4 architecture 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
Model 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.
Training 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.
Stage 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.
Stage 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-Bridge and Megatron-LM.
Stage 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
Stage 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.
Input
- 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
Output
- 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.
Software 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.
Model Version(s)
- v1.0 - GA
Quick Start Guide
Serving
vLLM
To deploy the Nemotron Labs 3 Puzzle NVFP4 checkpoint on NVIDIA Blackwell GPUs, 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 on vLLM v0.20.0.
- NVIDIA recommends setting
tpto2or4. - 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.
API 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)
Transformers
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, add enable_thinking=False to apply_chat_template(). By default, enable_thinking is set to True.
tokenized_chat = tokenizer.apply_chat_template(
messages,
tokenize=True,
enable_thinking=False,
add_generation_prompt=True,
return_tensors="pt"
).to(model.device)
Training and Evaluation Datasets
Training
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.
Additional 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 the Nemotron-3-Nano training 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 the Nemotron-3-Super RL pipeline, including single-step tool-use comparison data and end-to-end sandbox RL environments.
Click to explore the full dataset catalogue used for trainingBase Pre-Training Corpus (Nemotron 3 Foundation)
The foundation of the model is trained on the Nemotron-3-Nano corpus, comprising the following collections:
| Dataset Collection | Token Counts | Description |
|---|---|---|
| Nemotron-CC-v2 & v2.1 | 9.13T | A 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-v1 | 427.9B | High-quality code tokens extracted from Common Crawl using the Lynx + LLM pipeline to preserve structure and equations. |
| Nemotron-Pretraining-Code-v1 & v2 | 1.09T | Curated GitHub code references with multi-stage filtering, deduplication, and large-scale synthetic code data. |
| Nemotron-CC-Math-v1 | 133.3B | High-quality math pre-training dataset preserving LaTeX formatting and mathematical structures. |
| Nemotron-Pretraining-Specialized-v1 | 336.4B | Synthetic datasets targeting specialized domains such as STEM reasoning and scientific coding. |
Public Datasets
Sign up to read complete case studies, access detailed metrics, and unlock all use cases.
Sign up to read complete case studies, access detailed metrics, and unlock all use cases.