We introduce Nemotron 3 Ultra, a 550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention language model. We pre-trained Nemotron 3 Ultra on 20 trillion text tokens, then extended the context length to 1M tokens, and post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD). Nemotron 3 Ultra is our most capable model yet, employing multiple key technologies - LatentMoE, Multi Token Prediction (MTP), NVFP4 pre-training, multi-environment RLVR, MOPD, and reasoning budget control. Nemotron 3 Ultra achieves up to ~6x higher inference throughput as compared to state-of-the-art publicly available LLMs while attaining on-par accuracy. The state-of-the-art accuracy, high inference throughput, and 1M token context length make Nemotron 3 Ultra ideal for long-running autonomous agentic tasks. We open-source the base, post-trained, and quantized checkpoints, along with the training data and recipe on HuggingFace.

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3× faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20

We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.

Extending the context window of large language models (LLMs) is getting popular recently, while the solution of augmenting LLMs with retrieval has existed for years. The natural questions are: i) Retrieval-augmentation versus long context window, which one is better for downstream tasks? ii) Can both methods be combined to get the best of both worlds? In this work, we answer these questions by studying both solutions using two state-of-the-art pretrained LLMs, i.e., a proprietary 43B GPT and Llama2-70B. Perhaps surprisingly, we find that LLM with 4K context window using simple retrieval-augmentation at generation can achieve comparable performance to finetuned LLM with 16K context window via positional interpolation on long context tasks, while taking much less computation. More importantly, we demonstrate that retrieval can significantly improve the performance of LLMs regardless of their extended context window sizes. Our best model, retrieval-augmented Llama2-70B with 32K context window, outperforms GPT-3.5-turbo-16k and Davinci003 in terms of average score on nine long context tasks including question answering, query-based summarization, and in-context few-shot learning tasks. It also outperforms its non-retrieval Llama2-70B-32k baseline by a margin, while being much faster at generation. Our study provides general insights on the choice of retrieval-augmentation versus long context extension of LLM for practitioners.

Pretraining auto-regressive large language models (LLMs) with retrieval demonstrates better perplexity and factual accuracy by leveraging external databases. However, the size of existing pretrained retrieval-augmented LLM is still limited (e.g., Retro has 7.5B parameters), which limits the effectiveness of instruction tuning and zero-shot generalization. In this work, we introduce Retro 48B, the largest LLM pretrained with retrieval. Specifically, we continue to pretrain a 43B GPT model on additional 100 billion tokens using the Retro augmentation method by retrieving from 1.2 trillion tokens. Notably, the obtained foundation model, Retro 48B, largely outperforms the counterpart GPT 43B trained on 1.2T tokens in terms of perplexity with only 2.58% additional GPU hours, demonstrating the significant scaling potential of the method. After instruction tuning on Retro, InstructRetro demonstrates significant improvement over the instruction-tuned GPT on a wide range of zero-shot tasks. Specifically, the average improvement of InstructRetro is 7% over its GPT counterpart across 8 short-form QA and reading comprehension tasks, 10% over GPT across 4 challenging long-form QA tasks, and 16% over GPT across 3 summarization tasks. Surprisingly, we find that one can ablate the encoder from InstructRetro architecture and directly use its decoder backbone, while achieving comparable results. Our results highlight the promising direction to obtain a better GPT decoder through continued pretraining with retrieval before instruction tuning. Our code and checkpoints are publicly available at: https://huggingface.co/nvidia/retro-48b-instruct-4k.

Large decoder-only language models (LMs) can be largely improved in terms of perplexity by retrieval (e.g., RETRO), but its impact on text generation quality and downstream task accuracy is unclear. Thus, it is still an open question: shall we pretrain large autoregressive LMs with retrieval? To answer it, we perform a comprehensive study on a scalable pre-trained retrieval-augmented LM (i.e., RETRO) compared with standard GPT and retrieval-augmented GPT incorporated at fine-tuning or inference stages. We first provide the recipe to reproduce RETRO up to 9.5B parameters while retrieving a text corpus with 330B tokens. Based on that, we have the following novel findings: i) RETRO outperforms GPT on text generation with much less degeneration (i.e., repetition), moderately higher factual accuracy, and slightly lower toxicity with a nontoxic retrieval database. ii) On the LM Evaluation Harness benchmark, RETRO largely outperforms GPT on knowledge-intensive tasks, but is on par with GPT on other tasks. Furthermore, we introduce a simple variant of the model, RETRO++, which largely improves open-domain QA results of original RETRO (e.g., EM score +8.6 on Natural Question) and significantly outperforms retrieval-augmented GPT in both fine-tuning and zero-shot evaluation settings. Our findings highlight the promising direction of pretraining autoregressive LMs with retrieval as future foundation models. We release our code and model at: https://github.com/NVIDIA/Megatron-LM/blob/main/tools/retro/README.md

Training large transformer models is one of the most important computational challenges of modern AI. In this paper, we show how to significantly accelerate training of large transformer models by reducing activation recomputation. Activation recomputation is commonly used to work around memory capacity constraints. Rather than storing activations for backpropagation, they are traditionally recomputed, which saves memory but adds redundant compute. In this work, we show most of this redundant compute is unnecessary because we can reduce memory consumption sufficiently without it. We present two novel yet very simple techniques: sequence parallelism and selective activation recomputation. In conjunction with tensor parallelism, these techniques almost eliminate the need to recompute activations. We evaluate our approach on language models up to one trillion parameters in scale and show that our method reduces activation memory by 5x, while reducing execution time overhead from activation recomputation by over 90%. For example, when training a 530B parameter GPT-3 style model on 2240 NVIDIA A100 GPUs, we achieve a Model Flops Utilization of 54.2%, which is 29% faster than the 42.1% we achieve using recomputation. Our implementation will be available in both Megatron-LM and NeMo-Megatron.

Approximate computing is a very promising design paradigm for crossing the CPU power wall, primarily driven by the potential to sacrifice output quality for significant gains in performance, energy, and fault tolerance. Unfortunately, existing solutions have primarily either focused on new programming models, or new hardware designs, leaving significant room between these two ends for software-based optimizations. To fill this void, additional efforts should target the compilation and runtime stages, which have a critical impact on controlling the interactions of the many approximate subcomputations to form a well-optimized application.This paper presents EMEURO, a neural-network (NN) based emulation and acceleration platform. By restructuring algorithms to have the same data flow as a NN, EMEURO is able to achieve significant speedup across several domains with minimal design effort. EMEURO uses novel NN-based approximate computing techniques, including methods for efficiently searching the high-dimension subroutine space, and fine-grain control of error during runtime. EMEURO is able to achieve 7x-109x maximum speedup over the original algorithm with 0.1%-10% approximation error.

As datasets continue to grow, neural network (NN) applications are becoming increasingly limited by both the amount of available computational power and the ease of developing high-performance applications. Researchers often must have expert systems knowledge to make their algorithms run efficiently. Although available computing power increases rapidly each year, algorithm efficiency is not able to keep pace due to the use of general purpose compilers, which are not able to fully optimize specialized application domains. Within the domain of NNs, we have the added knowledge that network architecture remains constant during training, meaning the architecture's data structure can be statically optimized by a compiler. In this paper, we present SONNC, a compiler for NNs that utilizes static analysis to generate optimized parallel code. We show that SONNC's use of static optimizations make it able to outperform hand-optimized C++ code by up to 7.8X, and MATLAB code by up to 24X. Additionally, we show that use of SONNC significantly reduces code complexity when using structurally sparse networks.

Restricted Boltzmann Machines (RBMs) - the building block for newly popular Deep Belief Networks (DBNs) - are a promising new tool for machine learning practitioners. However, future research in applications of DBNs is hampered by the considerable computation that training requires. In this paper, we describe a novel architecture and FPGA implementation that accelerates the training of general RBMs in a scalable manner, with the goal of producing a system that machine learning researchers can use to investigate ever-larger networks.Our design uses a highly efficient, fully-pipelined architecture based on 16-bit arithmetic for performing RBM training on an FPGA. We show that only 16-bit arithmetic precision is necessary, and we consequently use embedded hardware multiply-and-add (MADD) units. We present performance results to show that a speedup of 25-30X can be achieved over an optimized software implementation on a high-end CPU.

This paper covers the implementation and testing of a Deep Belief Network (DBN) for the purposes of document classification. Document classification is important because of the increasing need for document organization, particularly journal articles and online information. Many popular uses of document classification are for email spam filtering, topic-specific search engine (where we only want to search documents that fall under a certain topic), or sentiment detection (to determine whether a document has more of a positive or negative tone) [1]. DBNs using a mostly unsupervised clustering algorithm for training are implemented. The clustering algorithm groups documents together that fall under the same category. Once the documents are grouped, a supervised learning algorithm (SVM in this case) is used to “assign” labels to the groups. More information and a background on DBNs is given in the following section. Testing of the DBN is primarily performed along three axes: 1) the number of hidden neurons in each layer of the DBN (which determines a “shape” for the overall network), 2) the number of layers in the DBN, and 3) the number of iterations that are used to train each layer of the DBN. Additionally, classification results of the DBN are compared with results obtained from using support vector machine (SVM) and Naïve Bayes (NB) classifiers. Some final experiments investigate how vocabulary size and word preprocessing affects performance. Preliminary results indicate that the DBN, given the implementations used in this paper, is not a viable alternative to other forms of document classifiers. The DBN has a significantly longer training time, and no matter how many training iterations are used, the DBN appears to have worse accuracy than either the SVM or NB classifiers. A number of suggestions for improvement are given in the final section, at least one of which should greatly increase the performance of the DBN.

This paper describes a Java package for automatically extending WordNet and other semantic lexicons. Extending these semantic lexicons by traditional means of hand labeling word relationships is an expensive and laborious process. We used machine learning techniques to automatically extract words with a given relationship from a text corpus. The package is made to be flexible, allowing for various modules, such as new classifiers and semantic lexicons, to be “plugged-in.” The power of the package comes from its ability to seamlessly integrate the Stanford Parser with WordNet. Results obtained for various tests, particularly those done for a Naïve Bayes classifier, are promising.