Llama-3.3-Nemotron-Super-49B-v1.5 Overview

Description:

Llama-3.3-Nemotron-Super-49B-v1.5 is a significantly upgraded version of Llama-3.3-Nemotron-Super-49B-v1 and is a large language model (LLM) which is a derivative of Meta Llama-3.3-70B-Instruct (AKA the reference model). It is a reasoning model that is post trained for reasoning, human chat preferences, and agentic tasks, such as Retrieval-Augmented Generation (RAG) and tool calling. The model supports a context length of 128K tokens.

Llama-3.3-Nemotron-Super-49B-v1.5 is a model which offers a great tradeoff between model accuracy and efficiency. Efficiency (throughput) directly translates to savings. Using a novel Neural Architecture Search (NAS) approach, we greatly reduce the model’s memory footprint, enabling larger workloads, as well as fitting the model on a single GPU at high workloads (H200). This NAS approach enables the selection of a desired point in the accuracy-efficiency tradeoff. For more information on the NAS approach, please refer to this paper.

The model underwent a multi-phase post-training process to enhance both its reasoning and non-reasoning capabilities. This includes a supervised fine-tuning stage for Math, Code, Science, and Tool Calling. Additionally, the model went through multiple stages of Reinforcement Learning (RL) including Reward-aware Preference Optimization (RPO) for chat, Reinforcement Learning with Verifiable Rewards (RLVR) for reasoning, and iterative Direct Preference Optimization (DPO) for Tool Calling capability enhancements. The final checkpoint was achieved after merging several RL and DPO checkpoints.

This model is part of the Llama Nemotron Collection. You can find the other model(s) in this family here:

This model is ready for commercial use.

License/Terms of Use

GOVERNING TERMS: The NIM container is governed by the NVIDIA Software License Agreement and the Product-Specific Terms for NVIDIA AI Products; and the use of this model is governed by the NVIDIA Open Model License.

Additional Information: Llama 3.3 Community License Agreement. Built with Llama.

Get Help

NVIDIA Developer Community Forum

Get access to community knowledge base articles and support cases (https://forums.developer.nvidia.com/)

Enterprise Support

Get access to knowledge base articles and support cases or submit a ticket.\nYou are responsible for ensuring that your use of NVIDIA provided models complies with all applicable laws.

Deployment Geography:

Global

Use Case:

Llama-3.3-Nemotron-Super-49B-v1.5 is a general purpose reasoning and chat model intended to be used in English and coding languages. Other non-English languages (German, French, Italian, Portuguese, Hindi, Spanish, and Thai) are also supported. Developers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. Also suitable for typical instruction-following tasks.

Release Date:

Hugging Face 7/25/2025 via Llama-3_3-Nemotron-Super-49B-v1_5
Build.NVIDIA.com 7/25/2025 Llama-3_3-Nemotron-Super-49B-v1_5

References

Model Architecture

Architecture Type: Dense decoder-only Transformer model

Network Architecture: Llama 3.3 70B Instruct, customized through Neural Architecture Search (NAS)

The model is a derivative of Meta’s Llama-3.3-70B-Instruct, using Neural Architecture Search (NAS). The NAS algorithm results in non-standard and non-repetitive blocks. This includes the following:

Skip attention: In some blocks, the attention is skipped entirely, or replaced with a single linear layer. Variable FFN: The expansion/compression ratio in the FFN layer is different between blocks.

We utilize a block-wise distillation of the reference model, where for each block we create multiple variants providing different tradeoffs of quality vs. computational complexity, discussed in more depth below. We then search over the blocks to create a model which meets the required throughput and memory (optimized for a single H100-80GB GPU) while minimizing the quality degradation. The model then undergoes knowledge distillation (KD), with a focus on English single and multi-turn chat use cases. The KD step included 40 billion tokens consisting of a mixture of 3 datasets - FineWeb, Buzz-V1.2, and Dolma.

Intended Use
\nLlama-3.3-Nemotron-Super-49B-v1.5 is a general purpose reasoning and chat model intended to be used in English and coding languages. Other non-English languages (German, French, Italian, Portuguese, Hindi, Spanish, and Thai) are also supported.

Input

Input Type(s): Text
Input Format: String
Input Parameters: One-Dimensional (1D)
Other Properties Related to Input: Context length up to 131,072 tokens

Output

Output Type(s): Text
Output Format: String
Output Parameters: One-Dimensional (1D)
Other Properties Related to Output: Context length up to 131,072 tokens

Our AI models are designed and/or 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: vLLM, TensorRT
Supported Hardware Microarchitecture Compatibility:\n
- NVIDIA Ampere
- NVIDIA Hopper
\n
Preferred 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:

Llama-3.3-Nemotron-Super-49B-v1.5-1.12.0
\nLlama-3.3-Nemotron-Super-49B-v1.5-1.13.1
\nLlama-3.3-Nemotron-Super-49B-v1.5-1.14.0
\nLlama-3.3-Nemotron-Super-49B-v1.5-1.15.0
\nLlama-3.3-Nemotron-Super-49B-v1.5-1.15.5

Quick Start and Usage Recommendations:

By default (empty system prompt), the model will respond in reasoning ON mode. Setting /no_think in the system prompt will enable reasoning OFF mode.
We recommend setting temperature to 0.6 and Top P to 0.95 for Reasoning ON mode.
We recommend using greedy decoding for Reasoning OFF mode.

You can try this model out through the preview API, using this link: Llama-3_3-Nemotron-Super-49B-v1_5.

Use It with vLLM

pip install vllm==0.9.2

An example on how to serve with vLLM:

$ python3 -m vllm.entrypoints.openai.api_server \\\n  --model \"nvidia/Llama-3_3-Nemotron-Super-49B-v1_5\" \\\n  --trust-remote-code \\\n  --seed=1 \\\n  --host=\"0.0.0.0\" \\\n  --port=5000 \\\n  --served-model-name \"Llama-3_3-Nemotron-Super-49B-v1_5\" \\\n  --tensor-parallel-size=8 \\\n  --max-model-len=65536 \\\n  --gpu-memory-utilization 0.95 \\\n  --enforce-eager\n

Running a vLLM Server with Tool-call Support

To enable tool calling usage with this model, we provide a tool parser in the repository. Here is an example on how to use it:

$ git clone https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5\n\n$ conda create -n vllm python=3.12 -y\n$ conda activate vllm\n$ pip install vllm==0.9.2\n\n$ python3 -m vllm.entrypoints.openai.api_server \\\n  --model Llama-3_3-Nemotron-Super-49B-v1_5 \\\n  --trust-remote-code \\\n  --seed=1 \\\n  --host=\"0.0.0.0\" \\\n  --port=5000 \\\n  --served-model-name \"Llama-3_3-Nemotron-Super-49B-v1_5\" \\\n  --tensor-parallel-size=8 \\\n  --max-model-len=65536 \\\n  --gpu-memory-utilization 0.95 \\\n  --enforce-eager \\\n  --enable-auto-tool-choice \\\n  --tool-parser-plugin \"Llama-3_3-Nemotron-Super-49B-v1_5/llama_nemotron_toolcall_parser_no_streaming.py\" \\\n  --tool-call-parser \"llama_nemotron_json\"\n

After launching a vLLM server, you can call the server with tool-call support using a Python script like below.

\nOkay, let's see. The user has a bill of $100 and wants to know the amount for an 18% tip. Hmm, I need to calculate the tip based on the bill total and the percentage. The tools provided include calculate_tip, which takes bill_total and tip_percentage as parameters. So the bill_total here is 100, and the tip_percentage is 18. I should call the calculate_tip function with these values. Wait, do I need to check if the parameters are integers? The bill is $100, which is an integer, and 18% is also an integer. So that fits the function's requirements. I don't need to convert any currency here because the user is asking about a tip in the same currency. So the correct tool to use is calculate_tip with those parameters.\n\n'''\nprint(completion.choices[0].message.tool_calls)\n'''\n[ChatCompletionMessageToolCall(id='chatcmpl-tool-e341c6954d2c48c2a0e9071c7bdefd8b', function=Function(arguments='{"bill_total": 100, "tip_percentage": 18}', name='calculate_tip'), type='function')]\n'''\n\">

from openai import OpenAI\nclient = OpenAI(\n    base_url=\"http://0.0.0.0:5000/v1\",\n    api_key=\"dummy\",\n)\ncompletion = client.chat.completions.create(\n    model=\"Llama-3_3-Nemotron-Super-49B-v1_5\",\n    messages=[\n        {\"role\": \"system\", \"content\": \"\"},\n        {\"role\": \"user\", \"content\": \"My bill is $100. What will be the amount for 18% tip?\"}\n    ],\n    tools=[\n        {\n            \"type\": \"function\",\n            \"function\": {\n                \"name\": \"calculate_tip\",\n                \"parameters\": {\n                    \"type\": \"object\",\n                    \"properties\": {\n                        \"bill_total\": {\n                            \"type\": \"integer\",\n                            \"description\": \"The total amount of the bill\"\n                        },\n                        \"tip_percentage\": {\n                            \"type\": \"integer\",\n                            \"description\": \"The percentage of tip to be applied\"\n                        }\n                    },\n                    \"required\": [\"bill_total\", \"tip_percentage\"]\n                }\n            }\n        },\n        {\n            \"type\": \"function\",\n            \"function\": {\n                \"name\": \"convert_currency\",\n                \"parameters\": {\n                    \"type\": \"object\",\n                    \"properties\": {\n                        \"amount\": {\n                            \"type\": \"integer\",\n                            \"description\": \"The amount to be converted\"\n                        },\n                        \"from_currency\": {\n                            \"type\": \"string\",\n                            \"description\": \"The currency code to convert from\"\n                        },\n                        \"to_currency\": {\n                            \"type\": \"string\",\n                            \"description\": \"The currency code to convert to\"\n                        }\n                    },\n                    \"required\": [\"from_currency\", \"amount\", \"to_currency\"]\n                }\n            }\n        }\n    ],\n    temperature=0.6,\n    top_p=0.95,\n    max_tokens=32768,\n    stream=False\n)\nprint(completion.choices[0].message.content)\n'''\n<think>\nOkay, let's see. The user has a bill of $100 and wants to know the amount for an 18% tip. Hmm, I need to calculate the tip based on the bill total and the percentage. The tools provided include calculate_tip, which takes bill_total and tip_percentage as parameters. So the bill_total here is 100, and the tip_percentage is 18. I should call the calculate_tip function with these values. Wait, do I need to check if the parameters are integers? The bill is $100, which is an integer, and 18% is also an integer. So that fits the function's requirements. I don't need to convert any currency here because the user is asking about a tip in the same currency. So the correct tool to use is calculate_tip with those parameters.\n</think>\n'''\nprint(completion.choices[0].message.tool_calls)\n'''\n[ChatCompletionMessageToolCall(id='chatcmpl-tool-e341c6954d2c48c2a0e9071c7bdefd8b', function=Function(arguments='{\"bill_total\": 100, \"tip_percentage\": 18}', name='calculate_tip'), type='function')]\n'''\n

Training, Testing, and Evaluation Datasets:

Training Dataset:

A large variety of training data was used for the knowledge distillation phase before the post-training pipeline, 3 of which included: FineWeb, Buzz-V1.2, and Dolma.

The data for the multi-stage post-training phases for improvements in Code, Math, and Reasoning is a compilation of SFT and RL data that supports improvements of math, code, general reasoning, and instruction following capabilities of the original Llama instruct model.

Prompts have been sourced from either public and open corpus or synthetically generated. Responses were synthetically generated by a variety of models, with some prompts containing responses for both reasoning on and off modes, to train the model to distinguish between two modes.

NVIDIA will be releasing the post-training dataset in the coming weeks.

Data Modality: Text

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

Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

Properties (Quantity, Dataset Descriptions, Sensor(s)):

Quantity: The knowledge distillation phase used 40 billion tokens. The underlying Dolma dataset contains approximately 3 trillion tokens, and FineWeb contains over 15 trillion tokens.

Dataset Descriptions: The training data is a composite of large-scale web text, synthetically generated conversational data for coding and instruction-following, and a diverse corpus of text from academic, literary, and encyclopedic sources. The model supports a context length of 128,000 tokens.

Testing Dataset:

Data Collection Method by dataset: Hybrid: Human, Synthetic

Labeling Method by dataset: Hybrid: Human, Automated

Properties (Quantity, Dataset Descriptions, Sensor(s)):

Quantity: The benchmarks contain thousands of test items, including:

MMLU: ~15,900 multiple-choice questions.
GPQA: 448 difficult multiple-choice questions.
HumanEval: 164 programming problems.
GSM-8K: ~8,500 grade-school math problems.
MATH: 12,500 competition math problems.
IF-Eval: Over 400 instruction-following prompts.

Dataset Descriptions: The evaluation suite is diverse, covering 57 academic and professional subjects (MMLU), expert-level reasoning (GPQA), Python code generation (HumanEval), mathematical problem-solving (GSM-8K, MATH), and the ability to follow precise instructions (IF-Eval).

Evaluation Dataset:

We used the datasets listed below to evaluate Llama-3.3-Nemotron-Super-49B-v1.5.

Data Collection Method by dataset: Hybrid: Human. Synthetic

Labeling Method by dataset: Hybrid: Human, Synthetic, Automatic

Evaluation Results
\nWe evaluate the model using temperature=0.6, top_p=0.95, and 64k sequence length. We run the benchmarks up to 16 times and average the scores to be more accurate.