Overview
Run NVIDIA Nemotron 3 Nano 30B A3B with high-performance inference. Deploy NVIDIA’s hybrid Mamba-Transformer MoE with 3.5B active parameters in a few clicks.
Explore additional models
Frequently
asked questions
Nemotron 3 Nano 30B A3B is good for high-volume RAG, chat, structured extraction, and the routine sub-tasks inside agent loops, where throughput matters more than peak reasoning. It is a 31.6B hybrid Mamba-Transformer mixture-of-experts with about 3.5B active parameters and a 262,144-token context, served in FP8.
Nemotron 3 Super 120B is the most capable Nemotron model on Crusoe Cloud for text reasoning, while Nemotron 3 Nano 30B A3B and Nemotron 3.5 Lightning 30B A3B are both about 3B-3.5B active per token and built for throughput. Lightning is the newer generation, with improved reasoning and instruction following. It is available for Serverless Inference, Self-Serve Deployments, and Serverless Fine-Tuning. Choose Nano models when you need to fine-tune a small, low-cost model, Lightning when you need more capacity to customize, and Super for reasoning depth.
The “A3B” in Nemotron 3 Nano 30B A3B stands for about 3 billion active parameters. The model is a mixture-of-experts with roughly 30B total parameters, but it routes each token through only about 3.5B of them. That means it holds 30B-class knowledge in memory while computing like a much smaller model, which keeps latency low and serving costs down at high request volumes.
Nemotron 3 Nano 30B A3B delivers 30B-class quality on reasoning, coding, and tool calling while activating only about 3.5B parameters per token, so it serves at small-model speed and cost. Its hybrid design keeps long-context inference efficient across the full 262,144-token window, which makes it a strong fit for retrieval-heavy and agentic workloads at scale.
A hybrid Mamba-Transformer model replaces most attention layers with Mamba state-space layers, which process long sequences in linear rather than quadratic time, and keeps a small number of attention layers for precise recall. In Nemotron 3 Nano 30B A3B, that design combined with mixture-of-experts routing lets a 30B-class model serve a 262,144-token context at small-model speed.
MemoryAlloy™ technology is a proprietary memory fabric in Crusoe’s inference engine that shares the KV cache across nodes, with cache-aware routing that sends each request to the node most likely to have its context ready. For Nemotron 3 Nano 30B A3B, repeated system prompts, retrieved documents, and agent histories skip redundant prefill, so time to first token stays low at high request volumes, and reused context is billed at the lower cached-input rate.


%204.png)
