NVIDIA Nemotron 3 Nano 30B
NVIDIA's open hybrid MoE model with Mamba-2 layers, configurable reasoning, and agentic tool use.
Overview
What is NVIDIA Nemotron 3 Nano 30B
NVIDIA Nemotron 3 Nano 30B is an open-weights hybrid MoE language model that combines Mamba-2 and Transformer attention layers. It activates 3.5 billion of its 30 billion parameters per token to deliver efficient reasoning, function calling, and long-context performance for agentic AI tasks.
Running it privately on Venice
On Venice, Nemotron 3 Nano runs privately with zero retention — your prompts are not stored or used for training. You get native tool use, web search, and structured JSON output alongside its configurable reasoning traces, all at a fraction of the cost of closed rivals. Because the weights are open, you maintain full sovereignty over the model if you choose to self-host later.
Assessment
Strengths and limitations
- Fully open weights, training data, and recipes: you can self-host or fine-tune without vendor lock-in.
- Hybrid MoE architecture (Mamba-2 + Attention) delivers high inference throughput with only 3.5B active parameters per token.
- Configurable reasoning traces let you show or hide chain-of-thought for harder prompts.
- Native tool use, web search, and structured JSON output on Venice make it agent-ready out of the box.
- Extremely low per-token cost on Venice compared to closed rivals.
- The 3.5B active parameter count can cap depth on highly complex reasoning versus larger closed models.
- On Venice the context window is 128K tokens, narrower than the 1M-token support available in native NVIDIA deployments.
- Not uncensored: it carries standard safety alignment that may refuse certain requests.
- Multilingual support is limited to six languages: English, German, Spanish, French, Italian, and Japanese.
Capabilities
What it supports
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
Specifications
Datasheet
- Maker
- NVIDIA
- Released
- December 2025
- Architecture
- Hybrid MoE (Mamba-2 + Attention layers)
- Parameters
- 30B total, 3.5B active per token
- Open weights
- Yes — fully open with training data and recipes
- Context window
- 128K tokens
- Max output
- 16.384K tokens
- Capabilities
- Function calling, Web search
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Jan 2026
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia-nemotron-3-nano-30b-a3b",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.07 per 1M input tokens and $0.30 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Open weights | Context window | Price (Venice) | Best for |
|---|---|---|---|---|
| NVIDIA Nemotron 3 Nano 30B | Yes | 128K tokens | $0.07 in · $0.30 out / 1M | Efficient agentic reasoning & tool use |
| DeepSeek V3.2 | Yes | 160K tokens | $0.33 in · $0.48 out / 1M | General-purpose long-context tasks |
| Google Gemma 4 31B Instruct | Yes | 256K tokens | $0.12 in · $0.36 out / 1M | Open long-context inference |
| Kimi K2.6 | Yes | 256K tokens | $0.75 in · $3.50 out / 1M | Deep reasoning & coding |
| Claude Sonnet 4.6 | No | 1M tokens | $3.60 in · $18 out / 1M | Enterprise accuracy & safety |
Open hybrid MoE with Mamba-2, configurable reasoning, and the lowest price among open-weights rivals.
Use cases
What it is good for
- 01Agentic workflows that combine function calling with web search and structured output.
- 02High-volume, cost-sensitive text generation and data extraction.
- 03Long-document summarization and analysis within the 128K context limit.
- 04Multi-step reasoning tasks where visible or hidden reasoning traces are useful.
- 05Self-hosted or fine-tuned deployments for data sovereignty.
Prompting
Getting better results
Enable reasoning traces in the chat template for math, logic, and hard coding prompts.
Use structured output mode when you need reliable JSON or schema-constrained responses.
Pair prompts with web search on Venice when the topic requires facts beyond the November 2025 training cutoff.
Place critical instructions at the start or end of long prompts to maximize attention within the 128K context window.
FAQ
Frequently asked questions
It is an open-weights hybrid MoE language model from NVIDIA that combines Mamba-2 and Transformer layers. It activates 3.5 billion of its 30 billion parameters per token and supports configurable reasoning traces, function calling, and web search.
Billed per token at $0.07 per 1M input tokens and $0.30 per 1M output tokens. There is no subscription required; you pay only for what you use.
Yes. NVIDIA releases the model with open weights, training data, and recipes, allowing self-hosting and fine-tuning. It is available on Hugging Face.
You can try it on Venice with a free account, which includes daily inference credits. Beyond that, usage is billed based on the per-token rates.
Yes. On Venice it supports function calling, structured JSON output, and web search, making it suitable for agentic workflows.
Yes. The model can be configured via its chat template to provide a final answer without intermediate reasoning traces, though accuracy on harder prompts may decrease slightly when reasoning is disabled.
Nemotron 3 Nano is smaller and cheaper, with a hybrid Mamba-2 architecture optimized for throughput and reasoning. DeepSeek V3.2 offers a larger 160K context window and broader general-purpose performance. Choose Nemotron for cost-efficient agentic tasks and DeepSeek for longer documents.
Claude Sonnet 4.6 is a closed model with a 1M context window and higher general reasoning accuracy. Nemotron 3 Nano is open-weights, far cheaper, and better for private, self-hostable agentic pipelines. Pick Claude for maximum accuracy on complex tasks; pick Nemotron for sovereignty and cost.
English, German, Spanish, French, Italian, and Japanese.
Run NVIDIA Nemotron 3 Nano 30B privately
No prompt logging. No data used for training.