---
library_name: transformers
license: other
license_name: nvidia-open-model-license
license_link: >-
https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/
pipeline_tag: text-generation
language:
- en
tags:
- nvidia
- Nemotron-Cascade
- reasoning
- general-purpose
- SFT
- RL
- pytorch
---
# Nemotron-Cascade-8B Intermediate ckpts
[](https://arxiv.org/abs/2512.13607)
[](https://huggingface.co/collections/nvidia/nemotron-cascade)
[](https://huggingface.co/collections/nvidia/nemotron-cascade)
[](https://huggingface.co/collections/nvidia/nemotron-cascade)
## Introduction
This repository releases the intermediate checkpoints produced during the development of [Nemotron-Cascade-8B](https://huggingface.co/nvidia/Nemotron-Cascade-8B). Nemotron-Cascade-8B is a general-purpose model trained using a sequential, domain-wise reinforcement learning pipeline, illustrated in the figure below.
We release checkpoints corresponding to each major stage of training:
- **Nemotron-Cascade-8B-SFT** (completed multi-stage SFT)
- **Nemotron-Cascade-8B-RLHF** (completed RLHF)
- **Nemotron-Cascade-8B-IFRL** (completed instruction following RL)
- **Nemotron-Cascade-8B-MathRL** (completed Math RL)
- **Nemotron-Cascade-8B-CodeRL** (completed Code RL)
The final model, [Nemotron-Cascade-8B](https://huggingface.co/nvidia/Nemotron-Cascade-8B), is obtained after the concluding SWE RL stage.
## Usage Recommendations
We recommend using RoPE scaling with the [YaRN](https://arxiv.org/abs/2309.00071) method to better support contexts longer than 32K. This can be enabled by updating the model’s `config.json` as shown below:
```json
{
...,
"rope_scaling": {
"rope_type": "yarn",
"factor": 2.0,
"original_max_position_embeddings": 32768
}
}
```
## Results
Same as [Nemotron-Cascade-8B](https://huggingface.co/nvidia/Nemotron-Cascade-8B), we use a maximum output length of 64K tokens for evaluation, with the temperature set to 0.6 and top-p to 0.95. We also apply RoPE scaling using the YaRN method with a scaling factor of 2.0.
| **Benchmark
Metric: Pass@1** | **Nemotron-
Cascade-8B-SFT** | **Nemotron-
Cascade-8B-RLHF** | **Nemotron-
Cascade-8B-IFRL** | **Nemotron-
Cascade-8B-MathRL** | **Nemotron-
Cascade-8B-CodeRL** | **Nemotron-
Cascade-8B** |
| :---- | :---: | :---: | :---: | :---: | :---: | :---: |
| ***Knowledge Reasoning*** |
| MMLU | 83.0 | 83.1 | 83.4 | 83.4 | 83.7 | 83.7 |
| MMLU Pro | 74.4 | 77.8 | 74.5 | 75.0 | 75.3 | 75.7 |
| GPQA-Diamond | 63.5 | 66.8 | 66.1 | 65.7 | 67.4 | 66.5 |
| ***Alignment*** |
| ArenaHard | 70.0 | 90.1 | 88.0 | 87.0 | 87.8 | 87.9 |
| IFEval (Strict Prompt) | 70.8 | 50.1 | 90.4 | 92.1 | 90.7 | 90.2 |
| IFBench | 21.2 | 24.5 | 40.5 | 40.4 | 38.1 | 40.8 |
| ***Math*** |
| AIME 2024 | 83.6 | 86.1 | 86.2 | 90.2 | 89.1 | 89.5 |
| AIME 2025 | 72.8 | 75.0 | 75.2 | 81.9 | 80.5 | 80.1 |
| ***Code*** |
| LCB v5 (08/24-02/25) | 59.2 | 70.2 | 70.2 | 70.6 | 75.3 | 74.3 |
| LCB v6 (08/24-05/25) | 56.7 | 67.2 | 66.7 | 67.4 | 71.5 | 71.1 |
| SWE Verified (Agentless) | 26.1 | 28.2 | 28.3 | 30.6 | 31.6 | 37.2 |
## Chat Template
All intermediate checkpoints use the same chat template as [Nemotron-Cascade-8B](https://huggingface.co/nvidia/Nemotron-Cascade-8B). Each is a unified model supporting both ***thinking*** and ***instruct*** (non-reasoning) modes. To switch between these two modes, simply append the `" /think"` (for ***thinking***) or the `" /no_think"` (for ***instruct***) tag to the end of the user input. See [Nemotron-Cascade-8B](https://huggingface.co/nvidia/Nemotron-Cascade-8B) for additional details.
## Release Date
Dec 19, 2025
## License
Your use of this model is governed by the [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/).
## Citation
```
@article{Nemotron_Cascade_Scaling_Cascaded_Reinforcement_Learning,
title={Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models},
author={Wang, Boxin and Lee, Chankyu and Lee, Nayeon and Lin, Sheng-Chieh and Dai, Wenliang and Chen, Yang and Chen, Yangyi and Yang, Zhuolin and Liu, Zihan and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei},
year={2025}
}
```