nvidia/Nemotron-Cascade-2-30B-A3B
Nemotron-Cascade-2-30B-A3B is an open 30 billion parameter Mixture-of-Experts (MoE) model from NVIDIA, featuring 3 billion activated parameters. Post-trained from Nemotron-3-Nano-30B-A3B-Base, it excels in reasoning and agentic capabilities, achieving gold medal performance in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI). This model supports both thinking and instruct modes and offers a 1M-token context length, making it suitable for complex problem-solving and advanced agentic tasks.
Loading preview...
Nemotron-Cascade-2-30B-A3B: Advanced Reasoning and Agentic MoE Model
NVIDIA's Nemotron-Cascade-2-30B-A3B is an open 30 billion parameter Mixture-of-Experts (MoE) model, utilizing 3 billion activated parameters. It is post-trained from the Nemotron-3-Nano-30B-A3B-Base and is specifically designed for strong reasoning and agentic capabilities.
Key Capabilities and Features
- Exceptional Reasoning: Achieves gold medal performance in the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI), demonstrating advanced problem-solving skills in mathematics and coding.
- Dual Operating Modes: Supports both 'thinking' and 'instruct' (non-thinking) modes, allowing for flexible interaction and detailed reasoning processes, with reasoning content enclosed in
<think>tags. - Extended Context Length: Features support for up to a 1 million-token context length, enabling the processing of very long inputs and complex multi-turn conversations.
- Agentic Functionality: Primarily supports OpenHands for agentic coding and Software Engineering (SWE) tasks, with specific tool-calling mechanisms.
- Optimized for Efficiency: As an MoE model, it balances performance with computational efficiency by activating a smaller subset of parameters per inference.
When to Use This Model
- Complex Mathematical and Coding Challenges: Ideal for applications requiring high-accuracy solutions in competitive programming, advanced mathematics, and code generation.
- Agentic Workflows: Suitable for developing intelligent agents that require robust reasoning and tool-use capabilities, particularly with OpenHands integration.
- Long-Context Applications: Beneficial for tasks that demand understanding and generating responses based on extensive textual information, thanks to its 1M-token context window.
- Instruction Following: Excels in instruction following and alignment, as indicated by strong performance on benchmarks like ArenaHard v2 and IFBench.