Measuring Open-Source Llama Nemotron Models on DeepResearch Bench
•
4
AI & ML interests
None defined yet.
Recent Activity
Articles
Open, Production-ready Enterprise Models
-
nvidia/Llama-3_3-Nemotron-Super-49B-v1_5
Text Generation • 50B • Updated • 7.89k • 152 -
nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8
Text Generation • 50B • Updated • 388 • 9 -
nvidia/Llama-3_1-Nemotron-Ultra-253B-v1
Text Generation • 253B • Updated • 143k • • 327 -
nvidia/Llama-3_3-Nemotron-Super-49B-v1
Text Generation • 50B • Updated • 21.1k • • 319
World Foundation Model for Future Prediction
Nemotron reward models. For use in RLHF pipelines and LLM-as-a-Judge
-
nvidia/Llama-3_3-Nemotron-Super-49B-GenRM
Text Generation • 50B • Updated • 209 • 13 -
nvidia/Llama-3_3-Nemotron-Super-49B-GenRM-Multilingual
Text Generation • 50B • Updated • 193 • 6 -
nvidia/Llama-3.3-Nemotron-70B-Reward
Text Generation • 71B • Updated • 54 • 2 -
nvidia/Llama-3.3-Nemotron-70B-Reward-Multilingual
Text Generation • 71B • Updated • 1.47k • 8
Math and Code reasoning model trained through reinforcement learning (RL)
Mamba-Transformer hybrid models
-
nvidia/Nemotron-H-47B-Reasoning-128K
Text Generation • 47B • Updated • 600 • 16 -
nvidia/Nemotron-H-8B-Reasoning-128K
Text Generation • 8B • Updated • 38.9k • 18 -
nvidia/Nemotron-H-8B-Reasoning-128K-FP8
Text Generation • 8B • Updated • 54 • 6 -
nvidia/Nemotron-H-47B-Reasoning-128K-FP8
Text Generation • 47B • Updated • 70 • 5
Multimodal Large Language Models for Detailed Localized Image and Video Captioning
Models and datasets from "AIMO-2 Winning Solution: Building State-of-the-Art Mathematical Reasoning Models with OpenMathReasoning dataset"
Reasoning data for supervised finetuning of LLMs to advance code generation and critique
Benchmarks for evaluating synthetic verifiers like test case generation and code reward models (as found in https://www.arxiv.org/abs/2502.13820).
Multimodal world understanding through reasoning
A suite of image and video tokenizers
A suite of image and video tokenizers
Collection of commercial-grade datasets for physical AI developers
The collection of Cosmos models
We are releasing math instruction models, math reward models, general instruction models, all training datasets, and a math reward benchmark.
Eagle 2 is a family of frontier vision-language models with vision-centric design. The model supports 4K HD input, long-context video, and grounding.
A series of Hybrid Small Language Models.
A family of frontier-class multimodal large language models (LLMs) that achieve state-of-the-art results on vision-language tasks and text-only tasks.
Nemotron-4: open models for Synthetic Data Generation (SDG). Includes Base, Instruct, and Reward models.
NeMo Parakeet ASR Models attain strong speech recognition accuracy while being efficient for inference. Available in CTC and RNN-Transducer variants.
-
nvidia/parakeet-rnnt-1.1b
Automatic Speech Recognition • Updated • 15.1k • 151 -
nvidia/parakeet-ctc-1.1b
Automatic Speech Recognition • 1B • Updated • 2.36k • 32 -
nvidia/parakeet-rnnt-0.6b
Automatic Speech Recognition • Updated • 421 • 10 -
nvidia/parakeet-ctc-0.6b
Automatic Speech Recognition • Updated • 7.44k • 14
InstructRetro is an autoregressive decoder-only language model (LM) with retrieval-augmented pretraining and instruction tuning.
A collection of models trained with Reinforcement Learning from Human Feedback (RLHF).
Llama3-ChatQA-1.5 models excel at conversational question answering (QA) and retrieval-augmented generation (RAG).
The Nemotron 3 8B Family of models is optimized for building production-ready generative AI applications for the enterprise.
MambaVision: A Hybrid Mamba-Transformer Vision Backbone. Includes both 1K and 21K pretrained models.
-
nvidia/MambaVision-L3-512-21K
Image Classification • 0.7B • Updated • 241 • 51 -
nvidia/MambaVision-L3-256-21K
Image Classification • 0.7B • Updated • 12 • 7 -
nvidia/MambaVision-L2-512-21K
Image Classification • 0.2B • Updated • 638 • 3 -
nvidia/MambaVision-L-21K
Image Classification • 0.2B • Updated • 139 • 4
A family of compressed models obtained via pruning and knowledge distillation
-
nvidia/Mistral-NeMo-Minitron-8B-Base
Text Generation • 8B • Updated • 8.73k • 176 -
nvidia/Mistral-NeMo-Minitron-8B-Instruct
Text Generation • 8B • Updated • 1.54k • 80 -
nvidia/Llama-3_1-Nemotron-51B-Instruct
Text Generation • 52B • Updated • 9.08k • 209 -
nvidia/Llama-3.1-Minitron-4B-Width-Base
Text Generation • 5B • Updated • 3.16k • 191
This is the collection that presents ChatQA-2, a suite of 128K long-context models, that also have exceptional RAG capabilities
Collection of models for OpenReasoning-Nemotron which are trained on 5M reasoning traces for Math, Code and Science.
-
nvidia/OpenReasoning-Nemotron-1.5B
Text Generation • 2B • Updated • 4.58k • 38 -
nvidia/OpenReasoning-Nemotron-7B
Text Generation • 8B • Updated • 4.69k • • 40 -
nvidia/OpenReasoning-Nemotron-14B
Text Generation • 15B • Updated • 829 • 33 -
nvidia/OpenReasoning-Nemotron-32B
Text Generation • 33B • Updated • 3.73k • • 109
Accelerated models for digital biology by the NVIDIA BioNeMo team. https://www.nvidia.com/en-us/clara/biopharma/
3D-Informed World-Consistent Video Generation with Precise Camera Control
A collection of generative models quantized and optimized with TensorRT Model Optimizer.
Joint video-text embedding for physical AI
Math reasoning models trained through reinforcement learning (RL)
Reasoning data for supervised finetuning of LLMs to advance data distillation for competitive coding
-
nvidia/OpenCodeReasoning
Viewer • Updated • 753k • 3.54k • 490 -
OpenCodeReasoning: Advancing Data Distillation for Competitive Coding
Paper • 2504.01943 • Published • 16 -
nvidia/OpenCodeReasoning-Nemotron-7B
Text Generation • 8B • Updated • 932 • • 37 -
nvidia/OpenCodeReasoning-Nemotron-14B
Text Generation • 15B • Updated • 790 • 18
Novel ITS approach for open-ended tasks - No. 1 on Arena Hard on 18 Mar 2025
Multimodal Conditional World Generation for World2World Transfer
World Foundation Model for Future Prediction
SOTA models on Arena Hard and RewardBench as of 1 Oct 2024.
-
nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
Text Generation • 71B • Updated • 98.3k • • 2.05k -
nvidia/Llama-3.1-Nemotron-70B-Reward-HF
71B • Updated • 1.1k • 88 -
nvidia/HelpSteer2
Viewer • Updated • 21.4k • 2.41k • 419 -
HelpSteer2-Preference: Complementing Ratings with Preferences
Paper • 2410.01257 • Published • 25
QLIP is a family of image tokenizers with SOTA reconstruction quality and zero-shot image understanding.
LLMs equipped with Dynamic Memory Compression to accelerate generation.
Essential datasets and models for content safety, topic-following, and security guardrails
-
nvidia/Aegis-AI-Content-Safety-Dataset-2.0
Viewer • Updated • 33.4k • 2.53k • 42 -
nvidia/llama-3.1-nemoguard-8b-topic-control
Text Classification • Updated • 1.25k • 15 -
nvidia/llama-3.1-nemoguard-8b-content-safety
Text Classification • Updated • 417 • 25 -
nvidia/CantTalkAboutThis-Topic-Control-Dataset
Viewer • Updated • 1.09k • 107 • 6
A series of Neural Audio Codecs
Collection of optimized ONNX model checkpoints for NVIDIA RTX GPUs
A collection of models and datasets introduced in "OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data"
A collection of models and datasets relating to SteerLM and HelpSteer.
A collection of multilingual and multitask speech to text models from NVIDIA NeMo 🐤
-
nvidia/canary-1b
Automatic Speech Recognition • Updated • 15k • 439 -
nvidia/canary-1b-flash
Automatic Speech Recognition • 0.8B • Updated • 141k • 233 -
nvidia/canary-180m-flash
Automatic Speech Recognition • Updated • 11.2k • 70 -
Training and Inference Efficiency of Encoder-Decoder Speech Models
Paper • 2503.05931 • Published • 3
A collection of models and datasets introduced in "OpenMathInstruct-1: A 1.8 Million Math Instruction Tuning Dataset"
NV-Embed is a generalist embedding model encompassing retrieval, reranking, classification, clustering, STS tasks.
A collection of Mamba-2-based research models with 8B parameters trained on 3.5T tokens for comparison with Transformers.
BigVGAN is a universal neural vocoder that generates audio waveform using mel spectrogram as input.
Enabling 4k resolution for VLMs, CVPR 2025, https://nvlabs.github.io/PS3/
A collection of Foundation Vision Models that combine multiple models (CLIP, DINOv2, SAM, etc.).
Classifier models that can be used in NeMo Curator for labelling/filtering datasets.
-
nvidia/domain-classifier
0.2B • Updated • 7.84k • 86 -
nvidia/quality-classifier-deberta
0.2B • Updated • 6.09k • 65 -
HuggingFaceFW/fineweb-edu-classifier
Text Classification • 0.1B • Updated • 23.5k • • 189 -
nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0
Text Classification • Updated • 266k • 24
Open, Production-ready Enterprise Models
-
nvidia/Llama-3_3-Nemotron-Super-49B-v1_5
Text Generation • 50B • Updated • 7.89k • 152 -
nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8
Text Generation • 50B • Updated • 388 • 9 -
nvidia/Llama-3_1-Nemotron-Ultra-253B-v1
Text Generation • 253B • Updated • 143k • • 327 -
nvidia/Llama-3_3-Nemotron-Super-49B-v1
Text Generation • 50B • Updated • 21.1k • • 319
Collection of models for OpenReasoning-Nemotron which are trained on 5M reasoning traces for Math, Code and Science.
-
nvidia/OpenReasoning-Nemotron-1.5B
Text Generation • 2B • Updated • 4.58k • 38 -
nvidia/OpenReasoning-Nemotron-7B
Text Generation • 8B • Updated • 4.69k • • 40 -
nvidia/OpenReasoning-Nemotron-14B
Text Generation • 15B • Updated • 829 • 33 -
nvidia/OpenReasoning-Nemotron-32B
Text Generation • 33B • Updated • 3.73k • • 109
World Foundation Model for Future Prediction
Accelerated models for digital biology by the NVIDIA BioNeMo team. https://www.nvidia.com/en-us/clara/biopharma/
Nemotron reward models. For use in RLHF pipelines and LLM-as-a-Judge
-
nvidia/Llama-3_3-Nemotron-Super-49B-GenRM
Text Generation • 50B • Updated • 209 • 13 -
nvidia/Llama-3_3-Nemotron-Super-49B-GenRM-Multilingual
Text Generation • 50B • Updated • 193 • 6 -
nvidia/Llama-3.3-Nemotron-70B-Reward
Text Generation • 71B • Updated • 54 • 2 -
nvidia/Llama-3.3-Nemotron-70B-Reward-Multilingual
Text Generation • 71B • Updated • 1.47k • 8
3D-Informed World-Consistent Video Generation with Precise Camera Control
Math and Code reasoning model trained through reinforcement learning (RL)
A collection of generative models quantized and optimized with TensorRT Model Optimizer.
Mamba-Transformer hybrid models
-
nvidia/Nemotron-H-47B-Reasoning-128K
Text Generation • 47B • Updated • 600 • 16 -
nvidia/Nemotron-H-8B-Reasoning-128K
Text Generation • 8B • Updated • 38.9k • 18 -
nvidia/Nemotron-H-8B-Reasoning-128K-FP8
Text Generation • 8B • Updated • 54 • 6 -
nvidia/Nemotron-H-47B-Reasoning-128K-FP8
Text Generation • 47B • Updated • 70 • 5
Joint video-text embedding for physical AI
Multimodal Large Language Models for Detailed Localized Image and Video Captioning
Math reasoning models trained through reinforcement learning (RL)
Models and datasets from "AIMO-2 Winning Solution: Building State-of-the-Art Mathematical Reasoning Models with OpenMathReasoning dataset"
Reasoning data for supervised finetuning of LLMs to advance data distillation for competitive coding
-
nvidia/OpenCodeReasoning
Viewer • Updated • 753k • 3.54k • 490 -
OpenCodeReasoning: Advancing Data Distillation for Competitive Coding
Paper • 2504.01943 • Published • 16 -
nvidia/OpenCodeReasoning-Nemotron-7B
Text Generation • 8B • Updated • 932 • • 37 -
nvidia/OpenCodeReasoning-Nemotron-14B
Text Generation • 15B • Updated • 790 • 18
Reasoning data for supervised finetuning of LLMs to advance code generation and critique
Novel ITS approach for open-ended tasks - No. 1 on Arena Hard on 18 Mar 2025
Benchmarks for evaluating synthetic verifiers like test case generation and code reward models (as found in https://www.arxiv.org/abs/2502.13820).
Multimodal world understanding through reasoning
Multimodal Conditional World Generation for World2World Transfer
A suite of image and video tokenizers
World Foundation Model for Future Prediction
A suite of image and video tokenizers
SOTA models on Arena Hard and RewardBench as of 1 Oct 2024.
-
nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
Text Generation • 71B • Updated • 98.3k • • 2.05k -
nvidia/Llama-3.1-Nemotron-70B-Reward-HF
71B • Updated • 1.1k • 88 -
nvidia/HelpSteer2
Viewer • Updated • 21.4k • 2.41k • 419 -
HelpSteer2-Preference: Complementing Ratings with Preferences
Paper • 2410.01257 • Published • 25
Collection of commercial-grade datasets for physical AI developers
QLIP is a family of image tokenizers with SOTA reconstruction quality and zero-shot image understanding.
The collection of Cosmos models
LLMs equipped with Dynamic Memory Compression to accelerate generation.
We are releasing math instruction models, math reward models, general instruction models, all training datasets, and a math reward benchmark.
Essential datasets and models for content safety, topic-following, and security guardrails
-
nvidia/Aegis-AI-Content-Safety-Dataset-2.0
Viewer • Updated • 33.4k • 2.53k • 42 -
nvidia/llama-3.1-nemoguard-8b-topic-control
Text Classification • Updated • 1.25k • 15 -
nvidia/llama-3.1-nemoguard-8b-content-safety
Text Classification • Updated • 417 • 25 -
nvidia/CantTalkAboutThis-Topic-Control-Dataset
Viewer • Updated • 1.09k • 107 • 6
Eagle 2 is a family of frontier vision-language models with vision-centric design. The model supports 4K HD input, long-context video, and grounding.
A series of Neural Audio Codecs
A series of Hybrid Small Language Models.
Collection of optimized ONNX model checkpoints for NVIDIA RTX GPUs
A family of frontier-class multimodal large language models (LLMs) that achieve state-of-the-art results on vision-language tasks and text-only tasks.
A collection of models and datasets introduced in "OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data"
Nemotron-4: open models for Synthetic Data Generation (SDG). Includes Base, Instruct, and Reward models.
A collection of models and datasets relating to SteerLM and HelpSteer.
NeMo Parakeet ASR Models attain strong speech recognition accuracy while being efficient for inference. Available in CTC and RNN-Transducer variants.
-
nvidia/parakeet-rnnt-1.1b
Automatic Speech Recognition • Updated • 15.1k • 151 -
nvidia/parakeet-ctc-1.1b
Automatic Speech Recognition • 1B • Updated • 2.36k • 32 -
nvidia/parakeet-rnnt-0.6b
Automatic Speech Recognition • Updated • 421 • 10 -
nvidia/parakeet-ctc-0.6b
Automatic Speech Recognition • Updated • 7.44k • 14
A collection of multilingual and multitask speech to text models from NVIDIA NeMo 🐤
-
nvidia/canary-1b
Automatic Speech Recognition • Updated • 15k • 439 -
nvidia/canary-1b-flash
Automatic Speech Recognition • 0.8B • Updated • 141k • 233 -
nvidia/canary-180m-flash
Automatic Speech Recognition • Updated • 11.2k • 70 -
Training and Inference Efficiency of Encoder-Decoder Speech Models
Paper • 2503.05931 • Published • 3
InstructRetro is an autoregressive decoder-only language model (LM) with retrieval-augmented pretraining and instruction tuning.
A collection of models and datasets introduced in "OpenMathInstruct-1: A 1.8 Million Math Instruction Tuning Dataset"
A collection of models trained with Reinforcement Learning from Human Feedback (RLHF).
NV-Embed is a generalist embedding model encompassing retrieval, reranking, classification, clustering, STS tasks.
Llama3-ChatQA-1.5 models excel at conversational question answering (QA) and retrieval-augmented generation (RAG).
A collection of Mamba-2-based research models with 8B parameters trained on 3.5T tokens for comparison with Transformers.
The Nemotron 3 8B Family of models is optimized for building production-ready generative AI applications for the enterprise.
BigVGAN is a universal neural vocoder that generates audio waveform using mel spectrogram as input.
MambaVision: A Hybrid Mamba-Transformer Vision Backbone. Includes both 1K and 21K pretrained models.
-
nvidia/MambaVision-L3-512-21K
Image Classification • 0.7B • Updated • 241 • 51 -
nvidia/MambaVision-L3-256-21K
Image Classification • 0.7B • Updated • 12 • 7 -
nvidia/MambaVision-L2-512-21K
Image Classification • 0.2B • Updated • 638 • 3 -
nvidia/MambaVision-L-21K
Image Classification • 0.2B • Updated • 139 • 4
Enabling 4k resolution for VLMs, CVPR 2025, https://nvlabs.github.io/PS3/
A family of compressed models obtained via pruning and knowledge distillation
-
nvidia/Mistral-NeMo-Minitron-8B-Base
Text Generation • 8B • Updated • 8.73k • 176 -
nvidia/Mistral-NeMo-Minitron-8B-Instruct
Text Generation • 8B • Updated • 1.54k • 80 -
nvidia/Llama-3_1-Nemotron-51B-Instruct
Text Generation • 52B • Updated • 9.08k • 209 -
nvidia/Llama-3.1-Minitron-4B-Width-Base
Text Generation • 5B • Updated • 3.16k • 191
A collection of Foundation Vision Models that combine multiple models (CLIP, DINOv2, SAM, etc.).
This is the collection that presents ChatQA-2, a suite of 128K long-context models, that also have exceptional RAG capabilities
Classifier models that can be used in NeMo Curator for labelling/filtering datasets.
-
nvidia/domain-classifier
0.2B • Updated • 7.84k • 86 -
nvidia/quality-classifier-deberta
0.2B • Updated • 6.09k • 65 -
HuggingFaceFW/fineweb-edu-classifier
Text Classification • 0.1B • Updated • 23.5k • • 189 -
nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0
Text Classification • Updated • 266k • 24