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-2
• reasoning
• general-purpose
• SFT
• RL
• heretic
• uncensored
• decensored
• abliterated

This is a decensored version of nvidia/Nemotron-Cascade-2-30B-A3B, made using Heretic v1.2.0

Abliteration parameters

Performance

Notes on changes to modeling_nemotron_h.py to resolve bugs in orginal code from NVIDIA

Nemotron-Cascade-2-30B-A3B ships with custom model code (modeling_nemotron_h.py) that has two bugs exposed by causal_conv1d 1.6.x. These are bugs in NVIDIA's model code, not in Heretic. Fix 1: Conv State Cache Dimension Mismatch File: modeling_nemotron_h.py (in the NemotronHCache.init method) What was wrong: The cache allocated conv_states tensors with intermediate_size (4096 = mamba_num_heads × mamba_head_dim), but the Mamba mixer's conv1d layer operates on the full projected dimension (6144 = intermediate_size + 2 × n_groups × ssm_state_size), which includes the B and C state channels concatenated with the hidden states. When causal_conv1d_update compared conv_state.shape[1] (4096) against weight.shape[0] (6144), it failed with "weight must have shape (dim, width)". The fix: Added a conv_dim variable matching the model's own formula from NemotronHMamba2Mixer.init, and used it for the cache allocation: intermediate_size = config.mamba_num_heads * config.mamba_head_dim conv_dim = intermediate_size + 2 * config.n_groups * config.ssm_state_size and torch.zeros(batch_size, conv_dim, conv_kernel_size, ...)

Fix 2: Prefill Sequence Length Guard File: modeling_nemotron_h.py (in cuda_kernels_forward) What was wrong: The single-token decode path was guarded by cache_position[0] > 0, which is true whenever the KV cache has been initialized — but modern transformers can still pass multi-token sequences after cache init (e.g., during the second forward call of generation). The causal_conv1d_update CUDA kernel expects a 2D (batch, dim) input for single-step updates, but received a 3D (batch, seq_len, dim) tensor, causing the shape check to fail. The fix: Added and hidden_states.shape[1] == 1 to the condition so the single-step CUDA path is only used when there's actually a single token: if cache_params is not None and cache_position is not None and cache_position[0] > 0 and hidden_states.shape[1] == 1

Nemotron-Cascade-2-30B-A3B

Technical Report SFT Dataset RL Dataset Models

Introduction

We're excited to introduce Nemotron-Cascade-2-30B-A3B, an open 30B MoE model with 3B activated parameters that delivers strong reasoning and agentic capabilities. It is post-trained from the Nemotron-3-Nano-30B-A3B-Base. Nemotron-Cascade-2-30B-A3B achieves gold medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). It operates in both thinking and instruct (non-thinking) modes.

Benchmark Results

Quick Start

• Nemotron-Cascade-2-30B-A3B follows the ChatML template and supports both thinking and instruct (non-thinking) modes. Reasoning content is enclosed within <think> and </think> tags. To activate the instruct (non-thinking) mode, we prepend <think></think> to the beginning of the assistant’s response.

• Nemotron-Cascade-2-30B-A3B does not currently support OpenCode; it primarily supports OpenHands for agentic coding and SWE tasks.

• To reduce the context length in a multi-turn conversation, when the previous user turn involves thinking mode, only the final summary of the model's output will be added to the conversation history.

• Note that we do not define a separate tool role for tool responses; instead, we place them under the user role and warp them with <tool_response> and </tool_response>.

• We recommend setting the sampling parameters to temperature = 1.0 and top_p = 0.95.

vLLM setup

Requires vLLM version >= 0.17.1. The following will create API endpoints at http://localhost:8000/v1:

• Standard version: Use the following command to create an API endpoint with a maximum context length of 262,144 tokens.

  vllm serve nvidia/Nemotron-Cascade-2-30B-A3B --port 8000 --tensor-parallel-size 1 --gpu-memory-utilization 0.9 --max-model-len 262144 --reasoning-parser nemotron_v3 --mamba-ssm-cache-dtype float32 --port 8000 --trust_remote_code
  

• Tool Call: Use the following command to enable tool support.

  vllm serve nvidia/Nemotron-Cascade-2-30B-A3B --port 8000 --tensor-parallel-size 1 --gpu-memory-utilization 0.9 --max-model-len 262144 --reasoning-parser nemotron_v3 --mamba-ssm-cache-dtype float32 --port 8000 --trust_remote_code --enable-auto-tool-choice --tool-call-parser qwen3_coder
  

Chat Template

from transformers import AutoTokenizer

model_name = 'nvidia/Nemotron-Cascade-2-30B-A3B'
tokenizer = AutoTokenizer.from_pretrained(model_name)

'''
single-turn example
'''
messages = [
  {"role": "system", "content": "You are a helpful and harmless assistant.\n\nYou are not allowed to use any tools"},
  {"role": "user", "content": "calculate 1+1?"}
]

# thinking mode
prompt_thinking = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
# prompt_thinking = '<|im_start|>system\nYou are a helpful and harmless assistant.\n\nYou are not allowed to use any tools<|im_end|>\n<|im_start|>user\ncalculate 1+1?<|im_end|>\n<|im_start|>assistant\n<think>\n'

# instruct mode
prompt_instruct = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
# prompt_instruct = '<|im_start|>system\nYou are a helpful and harmless assistant.\n\nYou are not allowed to use any tools<|im_end|>\n<|im_start|>user\ncalculate 1+1?<|im_end|>\n<|im_start|>assistant\n<think></think>'

'''
multi-turn example
'''
messages = [
    {"role": "system", "content": "You are a helpful and harmless assistant.\n\nYou are not allowed to use any tools"},
    {"role": "user", "content": "calculate 1+1?"},
    {"role": "assistant", "content": "<think>THINKING_CONTENT</think>\nTo calculate \\(1 + 1\\):\n\n1. **Identify the operation**: This is a basic addition problem involving two integers.\n2. **Perform the addition**:  \n   \\(1 + 1 = 2\\).\n\n**Result**: \\(\\boxed{2}\\)",},
    {"role": "user", "content": "what about 2+2"}
]

# thinking mode
prompt_thinking = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
# prompt_thinking = '<|im_start|>system\nYou are a helpful and harmless assistant.\n\nYou are not allowed to use any tools<|im_end|>\n<|im_start|>user\ncalculate 1+1?<|im_end|>\n<|im_start|>assistant\n<think></think>\nTo calculate \\(1 + 1\\):\n\n1. **Identify the operation**: This is a basic addition problem involving two integers.\n2. **Perform the addition**:  \n   \\(1 + 1 = 2\\).\n\n**Result**: \\(\\boxed{2}\\)<|im_end|>\n<|im_start|>user\nwhat about 2+2<|im_end|>\n<|im_start|>assistant\n<think>\n'

# instruct mode
prompt_instruct = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
# prompt_instruct = '<|im_start|>system\nYou are a helpful and harmless assistant.\n\nYou are not allowed to use any tools<|im_end|>\n<|im_start|>user\ncalculate 1+1?<|im_end|>\n<|im_start|>assistant\n<think></think>\nTo calculate \\(1 + 1\\):\n\n1. **Identify the operation**: This is a basic addition problem involving two integers.\n2. **Perform the addition**:  \n   \\(1 + 1 = 2\\).\n\n**Result**: \\(\\boxed{2}\\)<|im_end|>\n<|im_start|>user\nwhat about 2+2<|im_end|>\n<|im_start|>assistant\n<think></think>'

Python Tool Use

model_name = 'nvidia/Nemotron-Cascade-2-30B-A3B'
tokenizer = AutoTokenizer.from_pretrained(model_name)

SYSTEM_PROMPT = """# Tools

You have access to the following functions:

<tools>
<function>
<name>stateful_python_code_exec</name>
<description>Call this function to execute Python code in a stateful Jupyter notebook environment. Python will respond with the output of the execution or time out after 120.0 seconds.</description>
<parameters>
<parameter>
<name>code</name>
<type>string</type>
<description>Code to execute</description>
</parameter>
<required>["code"]</required>
</parameters>
</function>
</tools>

If you choose to call a function ONLY reply in the following format with NO suffix:

<tool_call>
<function=example_function_name>
<parameter=example_parameter_1>
value_1
</parameter>
<parameter=example_parameter_2>
This is the value for the second parameter
that can span
multiple lines
</parameter>
</function>
</tool_call>

<IMPORTANT>
Reminder:
- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags
- Required parameters MUST be specified
- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after
- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls
</IMPORTANT>"""

messages = [
  {"role": "system", "content": SYSTEM_PROMPT},
  {"role": "user", "content": "Solve the following math problem. Put your answer inside \\boxed{}.\n\nIn a school with 2008 students, each student is a member of certain committees. Each committee has at most 1004 members, and every two students are in at least one common committee. Determine the smallest possible number of committees in the school."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
print(prompt)

Agentic Usage

model_name = 'nvidia/Nemotron-Cascade-2-30B-A3B'
tokenizer = AutoTokenizer.from_pretrained(model_name)

SYSTEM_PROMPT = """You are a customer service agent that helps the user.  The policy that determines how you should respond to requests from users is described below between the <policy> and </policy> tags.

In each turn you can either:
- Send a message to the user.
- Make a tool call.
You cannot do both at the same time.

<policy>
_NEED_TO_ADD_POLICY_HERE_
</policy>

Try to be helpful and always follow the policy.

# Tools

You have access to the following functions:

<tools>
<function>
<name>_NEED_TO_ADD_FUNCTION_NAME_1_</name>
<description>_FUNCTION_DESCRIPTION_</description>
<parameters>
<parameter>
<name>_NEED_TO_ADD_PARAMETER_NAME_1_</name>
<type>_PARAMETER_TYPE_</type>
<description>_PARAMETER_DESCRIPTION_</description>
<title>_PARAMETER_TITLE_</title>
</parameter>
<parameter>
<name>_NEED_TO_ADD_PARAMETER_NAME_2_</name>
<type>_PARAMETER_TYPE_</type>
<description>_PARAMETER_DESCRIPTION_</description>
<title>_PARAMETER_TITLE_</title>
</parameter>
...... (_MORE_PARAMETERS_TO_ADD_)
<parameters>
</function>
...... (_MORE_FUNCTIONS_TO_ADD_)
</tools>
"""

messages = [
  {"role": "system", "content": SYSTEM_PROMPT},
  {"role": "user", "content": "Hello, I'm calling regarding my upcoming stay at your hotel. My guest ID is G90920 and booking ID is B11246 for a Deluxe room on June 5th. I'm traveling with three 6-month-old triplets and need to request three infant cribs for our room. It's currently 30 hours before check-in—could you please confirm if this is feasible and if there are quiet room options available for families with infants?"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
print(prompt)

Release Date

Mar 19, 2026

License

Your use of this model is governed by the NVIDIA Open Model License.

Citation

@article{Nemotron_Cascade_2,
  title={Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation},
  author={Yang, Zhuolin and Liu, Zihan and Chen, Yang and Dai, Wenliang and Wang, Boxin and Lin, Sheng-Chieh and Lee, Chankyu and Chen, Yangyi and Jiang, Dongfu and He, Jiafan and Pi, Renjie and Lam, Grace and Lee, Nayeon and Bukharin, Alexander and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei},
  year={2026}
}