davidnichols-ops/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-fork
NVIDIA Nemotron-3-Nano-30B-A3B-BF16 is a 30 billion parameter large language model developed by NVIDIA, featuring a hybrid Mixture-of-Experts (MoE) architecture with Mamba-2 and Attention layers. Designed for both reasoning and non-reasoning tasks, it can generate reasoning traces for higher accuracy on complex prompts. The model supports a maximum context length of 1M tokens and is optimized for agentic reasoning, general knowledge, and multilingual applications in English, German, Spanish, French, Italian, and Japanese.
Loading preview...
Model Overview
NVIDIA Nemotron-3-Nano-30B-A3B-BF16 is a 30 billion parameter large language model (LLM) developed by NVIDIA, featuring a unique hybrid Mixture-of-Experts (MoE) architecture. It combines 23 Mamba-2 and MoE layers with 6 Attention layers, utilizing 128 experts plus 1 shared expert per MoE layer, activating 6 experts per token. The model has 3.5 billion active parameters and supports an impressive 1 million token context length.
Key Capabilities
- Reasoning-First Approach: Designed to generate a reasoning trace before providing a final answer, which can be configured to enhance accuracy on complex tasks. Users can disable this for direct answers, though it may slightly reduce accuracy for reasoning-heavy prompts.
- Multilingual Support: Supports English, German, Spanish, French, Italian, and Japanese, with improved performance using Qwen.
- Agentic Reasoning: Excels in agentic tasks, demonstrating strong performance on benchmarks like AIME25 (with tools), MiniF2F, and SWE-Bench (OpenHands).
- Long Context Handling: Achieves high scores on long context benchmarks such as RULER-100@1M (86.3%).
Use Cases
This model is ideal for developers building AI Agent systems, chatbots, RAG systems, and other AI-powered applications requiring robust reasoning and instruction-following capabilities. Its configurable reasoning trace makes it adaptable for scenarios where either speed or accuracy is prioritized.