Skip to main content
Open SourceNVIDIA

NVIDIA Nemotron 3 Nano 30B A3B

NVIDIA Nemotron 3 Nano 30B A3B is a chat model from NVIDIA's Nemotron 3 family, positioned at the Nano size tier with a 30B/A3B parameter designation.

License Open Source
Input from
$0.060 / 1M tokens
on Liquid Inference · 1 provider
Output from
$0.239 / 1M tokens
on Liquid Inference

API Pricing

ProviderInput / 1MOutput / 1M
$0.060$0.239

Prices updated daily. Last check: Nov 19, 2026

Compare API pricing for every Nemotron model →

NVIDIA Nemotron 3 Nano 30B A3B pricing by provider

Input, output, and batch rates, plus alternatives, for one provider at a time.

Performance & Benchmarks

Source: Artificial Analysis →
Output Speed
304 t/s
Latency (TTFT)
500ms

Benchmarks measured Nov 2026. Scores are independent evaluations, not vendor-reported.

Model Details

General

Creator
NVIDIA
Modalities
Text

Capabilities

Open Source
Yes

Strengths & Limitations

Strengths

  • Measured output throughput of roughly 266 tokens per second in Artificial Analysis testing, fast enough for streaming chat UIs
  • Time to first token around 620 ms, keeping perceived response latency low for interactive use
  • A3B active-parameter design means only a small slice of the 30B parameter pool is computed per token, reducing serving cost relative to a dense 30B model
  • Nano tier sizing makes single-node or small-cluster GPU deployment practical rather than requiring multi-node inference
  • Part of NVIDIA's Nemotron 3 family, so it shares tooling and deployment paths with other Nemotron tiers
  • Suited to high-volume, repetitive workloads where per-request economics dominate

Limitations

  • We do not have a confirmed context window figure for this model in our database
  • No verified reasoning, coding, or knowledge benchmark scores are tracked for this entry, making capability comparisons difficult
  • Nano tier models generally offer less headroom on hard multi-step reasoning than larger tiers in the same family
  • Throughput and latency figures are provider- and load-dependent and will not match every endpoint listed here
  • Provider availability for Nemotron models is narrower than for the most widely hosted open-weight families

Key Features

•Chat/instruction-following text model
•Nemotron 3 family, Nano size tier
•30B total parameters with A3B (approximately 3B active) designation
•Measured ~266 output tokens/sec (Artificial Analysis)
•Measured ~620 ms time to first token (Artificial Analysis)
•Sparse active-parameter serving profile for lower per-token compute
•Multiple hosting providers compared in the pricing table on this page

About NVIDIA Nemotron 3 Nano 30B A3B

NVIDIA Nemotron 3 Nano 30B A3B is a text chat model released by NVIDIA as part of the Nemotron 3 line. Within that line it sits at the Nano tier, the smaller end of the family, and its name follows the increasingly common "total parameters / active parameters" convention: roughly 30 billion total parameters with an A3B (approximately 3 billion active) designation, indicating a sparse mixture-of-experts style architecture in which only a fraction of the network is engaged per token. NVIDIA has historically published Nemotron models as open-weight releases aimed at developers building on its own inference stack, though weight availability and licensing should be checked against NVIDIA's own model card for this specific release. The measured serving characteristics we track for this model come from Artificial Analysis: approximately 266 output tokens per second and a time to first token of about 620 ms. The high throughput figure is consistent with a small active-parameter count — the model only computes through a subset of its experts per token, so decode speed tends to resemble that of a much smaller dense model while the full parameter pool provides additional capacity. Note that both numbers are provider- and load-dependent and will differ across the endpoints listed on this page. We do not currently track a confirmed context window, modality set, or benchmark suite scores for this model; consult NVIDIA's model card for those details. In practice, a model at this size and active-parameter budget is typically deployed where per-request cost and latency matter more than maximum reasoning depth: high-volume assistant traffic, summarization and extraction pipelines, on-prem or single-node GPU serving, and agent loops that issue many short calls. Compared with larger Nemotron tiers, the Nano tier trades headroom on the hardest reasoning and coding tasks for faster generation and lighter hardware requirements.

Common Use Cases

Nemotron 3 Nano 30B A3B fits workloads where request volume and response speed matter more than peak reasoning capability: customer-facing chat assistants, document summarization, classification and data extraction pipelines, retrieval-augmented question answering over a fixed corpus, and agent loops that fire many short calls where a 620 ms time to first token and ~266 tokens/sec decode rate keep the overall loop responsive. The A3B active-parameter design makes it a candidate for teams running their own inference on NVIDIA hardware who want more parameter capacity than a small dense model without the serving cost of a full 30B dense forward pass. For tasks involving long multi-step derivations, difficult competitive coding, or deep research synthesis, a larger Nemotron tier or a dedicated reasoning model is usually the better starting point.

Frequently Asked Questions

How much does NVIDIA Nemotron 3 Nano 30B A3B cost to use?

Pricing depends on which provider hosts the model and on the pricing type — per-million-token API billing, hourly GPU rental for self-hosting, or a batch/committed rate. Rates change frequently and differ substantially between providers, so check the live pricing table on this page for current figures rather than relying on a static number.

What is this model best used for?

It is best suited to high-throughput chat and text-processing workloads: assistant traffic, summarization, extraction, classification, and short-turn agent calls. Its measured ~266 tokens/sec output rate and ~620 ms time to first token make it a reasonable fit for interactive, streaming interfaces where responsiveness matters.

What does "30B A3B" mean in the model name?

It denotes roughly 30 billion total parameters with approximately 3 billion active per token — a sparse mixture-of-experts style arrangement where only part of the network runs for any given token. The practical effect is decode speed closer to a small model while retaining a larger total parameter pool.

How does the Nano tier compare to larger Nemotron 3 models?

Nano is the smaller end of the Nemotron 3 line, optimized for speed and serving cost. Larger tiers in the family generally provide more headroom on complex reasoning, long-form coding, and knowledge-heavy tasks, at higher latency and cost per request. If your workload is mostly short, structured, or high-volume, the Nano tier is the more economical choice.

What context window does it support?

We do not currently have a verified context window figure for this model in our database. Check NVIDIA's model card or your chosen provider's API documentation, since some hosts also cap context below the model's native maximum.

Are the throughput numbers guaranteed?

No. The ~266 tokens/sec and ~620 ms time-to-first-token figures come from Artificial Analysis measurements and reflect one testing configuration. Real-world results vary by provider, hardware, batch size, prompt length, and current load.