Nvidia Nemotron 3 NVIDIA Open Source

Mamba-Transformer-MoE hybrid architecture, Apache 2.0 protocol open source

253B Total parameters
Apache 2.0 Open Source Agreement
Hybrid Architecture Mamba+Trans
Inference optimization Designed for inference

🟢 What is Nemotron 3?

Nemotron 3 is a new generation of open source large models released by Nvidia in December 2025. It uses innovative Mamba-Transformer-MoE A hybrid architecture that combines the efficient sequence modeling of Mamba, the powerful expressive power of Transformer, and the sparse computing advantages of MoE (Mixed Expert).

🏗️ Innovative architecture: Nemotron 3 integrates Mamba, Transformer and MoE architectures for the first time, greatly improving inference efficiency while maintaining strong performance. This hybrid architecture is a new direction in the development of large models.

Hybrid architecture diagram

Mamba + Transformer + MoE

🏗️ Architectural features

🐍 Mamba layer

State space model, linear complexity processing ultra-long sequences, efficient reasoning

🔄 Transformer layer

Attention mechanism handles complex dependencies and powerful context understanding

🎯 MoE sparse activation

Only some experts are activated in 253B parameters, greatly reducing calculation costs.

⚡ Inference optimization

Optimized for Nvidia GPU, natively supported by TensorRT-LLM

📊 Model specifications

Nemotron-3-253B-Instruct Ultimate version

The most powerful version, suitable for high-demand tasks such as complex reasoning, code generation, and long text understanding.

253B
Total parameters
128K
Context
Apache 2.0
Open Source Agreement

Nemotron-3-70B Balanced version

Balances performance and efficiency, suitable for most enterprise-level application scenarios.

70B
Total parameters
32K
Context
Efficient reasoning
Single card can run

Nemotron-3-8B Lightweight version

Suitable for edge deployment and resource-constrained scenarios, maintaining basic capabilities.

8B
Total parameters
8K
Context
Edge deployment
Consumer GPU

💻 Get started quickly

Loading Nemotron 3 using Hugging Face Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "nvidia/Nemotron-3-70B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

messages = [
    {"role": "user", "content": "解释一下 Mamba 架构的优势"}
]

input_ids = tokenizer.apply_chat_template(
    messages, 
    return_tensors="pt"
).to(model.device)

output = model.generate(input_ids, max_new_tokens=512)
print(tokenizer.decode(output[0], skip_special_tokens=True))
🚀 TensorRT-LLM optimization: Nvidia provides an official optimized version of TensorRT-LLM for significant inference acceleration on A100/H100 GPUs.

🎯 Typical application scenarios

🔬 Scientific research

Long text understanding, paper analysis, data processing

💻 Code generation

High-quality code completion, refactoring, and document generation

🤖 Agent Build

Build an intelligent agent with reasoning capabilities

🏢 Enterprise deployment

Apache 2.0 protocol, free for commercial use