nvidia/Nemotron-Cascade-2-30B-A3B

TEXT GENERATIONConcurrency Cost:2Model Size:30BQuant:FP8Ctx Length:32kPublished:Mar 18, 2026License:nvidia-open-model-licenseArchitecture:Transformer0.5K Open Weights Cold

Nemotron-Cascade-2-30B-A3B is an open 30 billion parameter Mixture-of-Experts (MoE) model from NVIDIA, with 3 billion activated parameters, designed for advanced reasoning and agentic capabilities. Post-trained from Nemotron-3-Nano-30B-A3B-Base, it achieves 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 features a 1M-token context length, making it highly suitable for complex problem-solving and agentic coding tasks.

Loading preview...

Nemotron-Cascade-2-30B-A3B Overview

Nemotron-Cascade-2-30B-A3B is a 30 billion parameter Mixture-of-Experts (MoE) model developed by NVIDIA, featuring 3 billion activated parameters. It is post-trained from the Nemotron-3-Nano-30B-A3B-Base and is distinguished by its strong reasoning and agentic capabilities.

Key Capabilities & Performance

  • Exceptional Reasoning: Achieves gold medal performance in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI), demonstrating advanced problem-solving skills.
  • Flexible Operation Modes: Supports both 'thinking' mode, where reasoning content is enclosed in <think> tags, and 'instruct' (non-thinking) mode.
  • Extended Context Length: Features an impressive 1M-token context length, enabling processing of very long inputs and multi-turn conversations.
  • Agentic Support: Primarily supports OpenHands for agentic coding and Software Engineering (SWE) tasks, with specific tool call parsing capabilities.
  • Benchmark Highlights: Outperforms several Nemotron-3 and Qwen3.5 models in benchmarks like IMO AnswerBench (79.3), LiveCodeBench v6 (87.2), and ArenaHard v2 (83.5 average).

Usage Recommendations

  • Chat Template: Follows the ChatML template, with specific handling for thinking and instruct modes.
  • Tool Use: Supports Python tool use and agentic workflows, with a defined system prompt structure for function calls.
  • Sampling Parameters: Recommended sampling parameters are temperature = 1.0 and top_p = 0.95.

Finetunes
3 models