NVIDIA Smashes Every Record: Achieves Gold-Medal Level With 20x Fewer Parameters
NVIDIA has just released an AI model that makes the impossible possible: Olympiad-level reasoning with only 3 billion active parameters. Nemotron-Cascade 2, launched on 19 March 2026,...
NVIDIA has just released an AI model that makes the impossible possible: Olympiad-level reasoning with only 3 billion active parameters.
Nemotron-Cascade 2, launched on 19 March 2026, is the second open-weight AI model ever to achieve Gold-Medal level performance in the International Mathematical Olympiad (IMO), the International Olympiad in Informatics (IOI), and the ICPC World Finals.
But here's the shocker: It does so with 20 times fewer parameters than its competitor DeepSeekV3.2-Speciale-671B-A37B.
What Does This Mean?
Conventional wisdom: Bigger models = better performance.
NVIDIA's evidence: No longer true.
Nemotron-Cascade 2 is a 30 billion parameter Mixture-of-Experts (MoE) model, but only 3 billion parameters are active during inference. For comparison:
- DeepSeekV3.2-Speciale: 671B total parameters, 37B active
- Nemotron-Cascade 2: 30B total parameters, 3B active
Yet both achieve Gold-Medal level on the most prestigious mathematics and programming competitions in the world.
Technical Breakthroughs
The NVIDIA team introduces two critical innovations:
1. Cascade RL (Reinforcement Learning)
Previously limited to specific tasks, now extended across the full spectrum of reasoning and agentic domains. The model teaches itself through a cascade of increasingly difficult challenges.
2. Multi-Domain On-Policy Distillation
Instead of one learning model, Nemotron-Cascade 2 distils knowledge from the strongest teacher models for each individual domain throughout the training process. This prevents "benchmark regression" (performance dropping when new skills are learned).
What Does This Mean for Norway?
Democratisation of AI:
Norwegian companies, universities and research institutions can now run Olympiad-level reasoning on consumer-grade GPUs. No need for billions in infrastructure.
The MedTech revolution:
Companies such as MediVox AI and Eir Tech can integrate advanced clinical reasoning into diagnostic tools without sending data to the cloud. Everything can be run locally with <4GB of GPU memory.
Education:
Norwegian students can gain access to an AI tutor that solves IMO problems at gold-medal level – free, open-source, and with no privacy concerns.
Availability
NVIDIA is following the trend from Meta and Anthropic: Full transparency.
- ✅ Model weights (open-weight)
- ✅ Training data
- ✅ Source code for reproduction
- ✅ Intermediate checkpoints
Everything is available on Hugging Face and NVIDIA NGC.
What Happens Now?
This is not the end, but the beginning. If Olympiad-level reasoning can be achieved with 3B parameters, what happens when we optimise further?
Next generation:
- 1B-parameter models that beat GPT-4 on specific domains
- Edge AI (mobile, IoT) with Olympiad-level capabilities
- Personalised agents that run on your phone without internet access
Nemotron-Cascade 2 proves that the AI revolution is not about size. It is about intelligence per parameter.
And that revolution has only just begun.
Sources:
- ArXiv: 2603.19220v1 (Yang et al., NVIDIA)
- NVIDIA Technical Blog
- Hugging Face Model Repository
Published: 21 March 2026
