# Nemotron 3 Nano 30B A3B

Thirty billion parameters of knowledge. Three billion parameters of compute per token. 262.1K tokens of context. Nemotron 3 Nano 30B A3B combines Mamba sequence modeling, Transformer attention, and MoE routing into a single hybrid architecture.

- **Model ID:** `nvidia/nemotron-3-nano-30b-a3b`
- **Type:** chat
- **Providers:** deepinfra
- **Context window:** 262,144
- **Maximum output tokens:** 262,144
- **Pricing:** $0.05/1M input tokens, $0.24/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/nemotron-3-nano-30b-a3b

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'nvidia/nemotron-3-nano-30b-a3b',
  prompt: 'Why is the sky blue?'
})
```

## About

NVIDIA announced Nemotron 3 Nano 30B A3B on December 15, 2025 as the first model in the Nemotron 3 family. The core idea is architectural efficiency at scale. 30B total parameters provide a broad knowledge base, but only 3B activate for any given token. This keeps inference cost and speed in the range of much smaller models.

Three layer types interleave throughout the architecture. Mamba-2 layers handle sequence processing with linear-time complexity. This makes the context window of 262.1K tokens feasible without the quadratic memory growth of pure attention. Transformer attention layers appear at strategic depths to maintain precise associative recall: the ability to pick out a specific fact from a large context. Mixture-of-experts (MoE) routing selects which expert parameters activate for each token, keeping compute proportional to the 3B active count rather than the full 30B.

Weights and recipes are available under the NVIDIA Open Model License. Deployment cookbooks for vLLM, SGLang, and TensorRT-LLM are also provided. Overview and techniques: https://deepinfra.com/nvidia/Nemotron-3-Nano-30B-A3B.

## What to consider

With a context window of 262.1K tokens, entire codebases or multi-document evidence sets fit in a single call. Plan context usage carefully. Filling the window is possible, but model the cost and latency implications ahead of time. Compare $0.05 and $0.24.

## When to use

### Best For

- Multi-agent systems running many concurrent lightweight agents where per-agent throughput matters
- Long-context tasks: holding entire codebases, extended session histories, or multi-document sets in one call
- Agentic workflows with multi-step tool calling

### Consider Alternatives When

- You need maximum reasoning depth. Nemotron 3 Super (120B/12B active) handles complex multi-agent planning
- You need vision-language capabilities. Nemotron Nano 12B v2 VL is the multimodal option
- A 128K context window is sufficient and the 262.1K tokens capacity goes unused

## Best for

- **Concurrent multi-agent systems:** Running many lightweight agents where per-agent throughput matters
- **Long-context tasks:** Holding entire codebases, extended session histories, or multi-document sets in one call
- **Agentic tool-calling workflows:** Multi-step pipelines with chained actions

## Consider alternatives

- **Maximum reasoning depth:** Nemotron 3 Super (120B/12B active) handles complex multi-agent planning
- **Vision-language tasks:** Nemotron Nano 12B v2 VL is the multimodal option
- **Smaller context needs:** A 128K context window is sufficient and the 262.1K tokens capacity goes unused
- **Compact dense reasoning:** Nemotron Nano 9B v2 targets a dense model profile

## Frequently asked questions

### Why does "30B total, 3B active" matter for inference cost?

You pay for compute proportional to the active parameters, not the total. Nemotron 3 Nano 30B A3B runs at speeds and costs closer to a 3B dense model but draws on 30B parameters of learned knowledge. The MoE routing mechanism selects the relevant subset per token.

### How does the Mamba architecture enable the context of 262.1K tokens?

Mamba layers process sequences with linear-time complexity rather than the quadratic scaling of standard attention. That makes it practical to hold 262.1K tokens in context without the memory explosion that would make pure-attention models infeasible at that length.

### How does Nemotron 3 Nano 30B A3B differ from Nemotron Nano 9B v2?

They use different architectures. Nemotron 3 Nano 30B A3B is a sparse MoE with 30B total/3B active parameters and a context window of 262.1K tokens. Nemotron Nano 9B v2 is a dense 9B model with a 128K-token context window. Choose Nemotron 3 Nano 30B A3B for throughput across multi-agent systems and Nano 9B v2 as a compact reasoning model.

### Where are hosted input and output prices listed?

Current pricing is shown on this page. AI Gateway routes across providers, and rates may vary by provider.

## Links

- [Model page](https://vercel.com/ai-gateway/models/nemotron-3-nano-30b-a3b)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://deepinfra.com/nvidia/Nemotron-3-Nano-30B-A3B)
- [Provider pricing](https://deepinfra.com/pricing)