Back to all models

Nemotron-3 Nano 30B A3B

Deploy Nemotron-3 Nano 30B A3B on a dedicated GPU from $0.45 per hour, billed per minute, on 1× RTX 4090. A sparse 30B hybrid Mamba-2/Transformer MoE that activates only about 3.5B parameters per token, with configurable reasoning, tool use and long-context agent serving.

Text30B paramsNVIDIA Nemotron Open Model LicenseDeploys in ~5 minSecure Cloud
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16

HexGrid Cloud price

$0.45/hr

1× RTX 4090

Parameters

30B

~3.5B active

Context

256K

1M extended

Architecture

Hybrid MoE

Mamba-2 + attention

Running Nemotron-3 Nano 30B A3B on HexGrid Cloud

What you get when you deploy with us, beyond the hourly rate.

Single tenant by default

The model runs on a GPU that is yours for the life of the instance. No shared endpoint, no queue behind other tenants.

Per-minute billing

You pay for the GPU, not the token. Stop the instance and billing stops with it.

OpenAI-compatible endpoint

Point an existing SDK at your instance by changing the base URL. No rewrite required.

Your weights, your data

Prompts and outputs stay inside your instance. Nothing is logged, sampled or used for training.

Nemotron-3 Nano 30B A3B GPU sizing and cost

What the model needs at each precision, and the cheapest HexGrid Cloud configuration that holds it.

Nemotron-3 Nano 30B A3B vRAM requirement and hourly cost by precision
PrecisionvRAM neededRuns onPrice
BF16
Recommended
72 GB5× RTX A4000$1.00/hr
FP8
36 GB1× RTX A6000$0.55/hr
INT4
18 GB1× RTX 4090$0.45/hr

NVIDIA publishes BF16, FP8 and NVFP4 deployment variants, but throughput depends on exact hardware and runtime. Do not substitute marketing throughput from a different GPU/configuration.

About Nemotron-3 Nano 30B A3B

What Nemotron-3 Nano 30B A3B is built for, and where it falls short.

Nemotron-3 Nano 30B A3B is NVIDIA's compact hybrid mixture-of-experts model for reasoning, chat and agent workloads. Although the model contains 30B parameters in total, NVIDIA states that roughly 3.5B are active during inference.

The network combines Mamba-2 sequence layers, sparse MoE feed-forward computation and a smaller number of grouped-query attention layers. This is intended to reduce active computation while retaining a larger parameter pool.

The same checkpoint supports reasoning and non-reasoning modes through the chat template, and NVIDIA documents tool calling, long-context serving and reasoning-budget controls.

What people run it for

Reasoning agents

Run agent steps with reasoning enabled, while controlling or disabling the reasoning trace when latency matters.

Coding agents

Use tool calling and code-oriented training for software and terminal workflows.

Long-context RAG

Process large retrieved contexts or accumulated agent state with 256K and optional extended-context serving.

High-throughput assistants

Use sparse activation to avoid paying dense 30B compute on every token.

Strengths

  • Only a fraction of the total parameters are active for each token
  • Unified reasoning and non-reasoning behavior from one checkpoint
  • Long-context operation documented at 256K with extension to 1M
  • Native tool-call serving examples for agentic applications
  • Hybrid Mamba-2 and attention architecture aimed at efficient sequence processing
  • Strong reasoning results including 89.1 on AIME25 without tools in NVIDIA's evaluation

Limitations

  • Custom hybrid architecture requires sufficiently recent inference-engine support
  • Sparse parameter count should not be compared directly with dense-model parameter counts as a measure of compute or memory
  • 1M-context operation requires substantially more memory than standard serving and explicit long-context configuration
  • Primary language emphasis is English and coding; NVIDIA officially lists only a limited set of additional natural languages
  • Distributed under NVIDIA's Nemotron licence rather than Apache 2.0

Quickstart

NVIDIA documents serving the BF16 checkpoint with vLLM using an OpenAI-compatible chat endpoint and dedicated reasoning/tool parsers.

pythonPOST /v1/chat/completions
from openai import OpenAI

client = OpenAI(
    base_url="https://<your-instance>.hexgrid.cloud/v1",
    api_key="<your-key>",
)

response = client.chat.completions.create(
    model="nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
    messages=[{"role": "user", "content": "Design a retry strategy for an unreliable API."}],
    max_tokens=4096,
)

print(response.choices[0].message.content)

NVIDIA's vLLM recipe uses the qwen3_coder tool parser and NVIDIA's nano_v3 reasoning parser. Reasoning can be disabled through chat-template kwargs.

Nemotron-3 Nano 30B A3B reported benchmarks

Scores published by the model authors. A starting point, not a guarantee on your own data.

  • MMLU-Pro78.3
  • AIME25 no tools89.1
  • GPQA no tools73
  • LiveCodeBench v668.3
  • BFCL v453.8
  • RULER-100 @ 256K92.9

Nemotron-3 Nano 30B A3B specifications

Architecture and serving details for Nemotron-3 Nano 30B A3B.

PublisherNVIDIAParameters30BLicenceNVIDIA Nemotron Open Model LicenseReleasedDecember 2025

Architecture

Type
Mamba2-Transformer Hybrid MoE
Total parameters
30B
Active parameters
~3.5B
Mamba-2 + MoE layers
23
Attention layers
6
Routed experts
128 per MoE layer
Experts activated
6 per token

Context

Standard serving example
262,144 tokens
Extended context
Up to 1M documented by NVIDIA
Input
Text
Output
Text

Serving

Engines
vLLM, TensorRT-LLM, SGLang
Reasoning
Configurable
Thinking budget
Supported
Tool calling
Supported

Nemotron-3 Nano 30B A3B frequently asked questions

Common questions about deploying Nemotron-3 Nano 30B A3B on HexGrid Cloud.

What does A3B mean?

It indicates that only roughly three billion parameters are active for each token even though the overall sparse model contains about 30 billion parameters. NVIDIA's current model card describes approximately 3.5B active parameters.

Can I disable reasoning?

Yes. NVIDIA documents enable_thinking=false through the chat template; reasoning is enabled by default in its serving example.

Does it support a 1M context?

NVIDIA documents 256K in its standard vLLM configuration and explains how to raise the maximum context to 1M. Long-context memory requirements remain deployment dependent.

Is it Apache 2.0?

No. It is governed by the NVIDIA Nemotron Open Model License.

Deploy Nemotron-3 Nano 30B A3B today

$0.45 per hour on 1× RTX 4090, billed per minute, never shared.