Key takeaways

• The Nemotron 3 Nano 30B A3B is a small, open-source Mixture of Experts (MoE) text-to-text language model (30B parameters) designed for building specialized, compute-efficient, and accurate agentic AI systems.
• The model demonstrates strong capabilities in multi-step arithmetic problem-solving, achieving 95.38% accuracy on the Grade School Math 8K dataset and excelling at interpreting percentages, fractions, and unit conversions.
• A significant limitation is its poor performance in scientific problem-solving (less than 1% pass rate on Humanity's Last Exam), struggling with complex logical conditions, advanced functions, and factual inaccuracies.
• Common failure modes include misinterpreting comparative language in math problems, logical errors in complex SQL queries (22.67% accuracy on BIRD-CRITIC dataset), and a lack of deep domain knowledge leading to factual inaccuracies and computational mistakes in scientific tasks.
• While the model can process and respond to complex multi-modal prompts (including CSV data and images), its accuracy is severely limited by frequent factual errors, logical inconsistencies, and computational mistakes, indicating a gap in deep domain knowledge and nuanced reasoning.

Technical specifications

Context length

256k tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~30B

Open weights

Yes

Licence

NVIDIA Open Model

Regions

USA

Vendor

Nvidia

Provider model key

nvidia/nemotron-3-nano-30b-a3b