01What Nemetron Cascade 2 Is and How It Works

Nemetron Cascade 2 is Nvidia's new open-source AI model that uses only 3 billion active parameters to match the performance of models with 600 billion or more. It is a mixture-of-experts (MoE) model, meaning it has 30 billion total parameters but only activates 3 billion per token—like a hospital sending each patient only to the specialist they need.

The model was trained using a method called Cascade Reinforcement Learning (Cascade RL), which teaches the AI one domain at a time. This prevents catastrophic forgetting, where improving at one skill hurts others. The result is a lightweight model that achieves gold-medal performance in the world's hardest math and coding competitions.

02Main Features of Nemetron Cascade 2

Nemetron Cascade 2 has several standout features. First, it is a mixture-of-experts model with 30 billion total parameters but only 3 billion active at any time. This cuts compute costs significantly while keeping performance high.

  • It operates in both instruct mode (direct answers) and thinking mode (step-by-step reasoning shown).
  • It was trained on 1.9 million Python reasoning traces, 1.3 million tool-calling samples, and 816,000 mathematical proof samples.
  • It includes 125,000 agentic samples, making it ideal for AI agents that take actions.
  • The model is open-source, free to download via Olama, and available on Hugging Face under Nvidia's open model license.

03How to Set Up Nemetron Cascade 2

Getting started with Nemetron Cascade 2 is straightforward. You can download and run it on your own hardware without any API costs or rate limits.

  1. Go to Olama's website or terminal and install Olama if you haven't already.
  2. Run the command to pull Nemetron Cascade 2: olama pull nemetron-cascade-2.
  3. Verify your hardware meets the minimum specs (30B total, 3B active—runs on many consumer GPUs).
  4. Launch the model locally and test it with a simple prompt like "Explain a math problem step by step."
  5. For advanced use, download the full checkpoint, SFT training data, and RL training data from Hugging Face under Nvidia's license.
  6. Customize the model for your own use case, such as agentic workflows or coding tasks.

04How to Use Nemetron Cascade 2 in Practice

In the episode, Julian showed how Nemetron Cascade 2 can handle complex reasoning and agentic tasks. Here are the key ways to use it based on the examples shown.

  • Math and coding: Ask the model to solve Olympiad-level math problems or write code. It shows its work in thinking mode, so you can catch mistakes.
  • Agentic workflows: Give it a goal like "Research competitors and draft a proposal." It will call tools, run code, read documents, and loop back if something fails.
  • Multi-agent systems: Run dozens or hundreds of agents simultaneously, all communicating and handing off tasks—thanks to high tokens-per-second throughput.
  • Offline use: Since it's open-source, you can run it completely offline, saving costs and keeping data private.

05Test Results and Comparisons Shown in the Episode

Julian shared specific benchmark results. Nemetron Cascade 2 won gold medals at the International Mathematical Olympiad (IMO), International Olympiad in Informatics (IOI), and ICPC World Finals—the three hardest math and coding competitions.

It outperformed the 122-billion-parameter Qwen 3.5 on complex math, coding, and agentic workflows. The only previous model to achieve gold across all three was DeepSeek v3.2 Speciale, which has 671 billion parameters. Nemetron Cascade 2 matched that with just 30 billion total and 3 billion active.

On the AIME25 math benchmark, the distillation approach reached teacher-level performance in 30 steps, while standard approaches scored lower after the same number of steps. A former IMO gold medalist personally verified the model's solutions.

06Tips, Limits, and Mistakes to Avoid with Nemetron Cascade 2

To get the most out of this model, follow these tips from the episode. First, always use thinking mode for tasks where you need to verify reasoning—this lets you catch errors.

Don't assume bigger is better. The model proves that size does not equal intelligence. Avoid judging models solely by their parameter count.

Also, remember that the model is open-source, so you are not locked into any API. Do not rely on closed APIs for critical workflows—if pricing changes, your systems break.

Finally, the training data is domain-specific. For tasks outside math, coding, or agentic work, you may need to fine-tune with your own data. The SFT and RL training data are available on Hugging Face to help with that.

07Who Nemetron Cascade 2 Is For and Why It Matters

Nemetron Cascade 2 is for developers, researchers, business owners, freelancers, and anyone building AI-powered tools. It matters because it shatters the assumption that you need a massive, expensive model to get frontier-level intelligence.

For business owners, this means you can run your own AI agents—handling customer emails, scheduling, research, and more—without paying high API fees. Freelancers and small agencies can now compete with larger teams. The open-source nature means the tools you use will get smarter and cheaper within weeks, putting you ahead if you learn to use them now.

As Julian said, the gap between people who understand AI tools and those who don't is growing every week. This model makes advanced reasoning accessible to anyone willing to try.