Loading page
NVIDIA
Nemotron 3 Nano Omni 30B A3B Reasoning
Nemotron 3 Nano Omni 30B A3B Reasoning is a compact reasoning-oriented model from NVIDIA's Nemotron 3 Nano line, measured at roughly 330 output tokens per second in third-party testing.
Input from
$0.300 / 1M tokens
across 1 provider
API Pricing
| Provider | Input / 1M | Output / 1M | Cached / 1M |
|---|---|---|---|
| $0.300 | $1.83 | $0.300 |
Prices updated daily. Last check: Sep 20, 2026
Nemotron 3 Nano Omni 30B A3B Reasoning pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Intelligence
10.3 / 100
Coding
13.8 / 100
Reasoning & Knowledge
- GPQA Diamond46.9%
- Humanity's Last Exam4.8%
Agentic & Tool Use
- Terminal-Bench Hard8.3%
- Terminal-Bench v2.16.7%
- τ²-bench45.3%
Instruction & Long Context
- IFBench63.2%
- Long-Context Reasoning39.7%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- NVIDIA
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
Strengths & Limitations
Strengths
- Measured at approximately 329.9 output tokens per second in Artificial Analysis testing, which shortens wall-clock time on verbose chain-of-thought responses
- Time to first token measured around 554 ms, suitable for interactive use rather than batch-only workloads
- Mixture-of-experts style naming (30B total / ~3B active) implies a small active parameter count per token, which typically reduces serving cost relative to dense models of comparable total size
- Reasoning-tuned variant produces explicit intermediate steps, useful for math, logic, and multi-constraint tasks where a direct-answer model tends to skip work
- Part of NVIDIA's Nemotron 3 family, so it shares tooling and deployment conventions with other Nemotron releases
- Small enough tier to be a candidate for self-hosted or on-prem serving where data cannot leave the environment
Limitations
- We do not currently track a confirmed context window for this variant — verify maximum input length with your provider
- Supported input modalities are not confirmed in our data despite the "Omni" naming; do not assume image or audio input without checking provider docs
- No accuracy benchmark scores are tracked for this model in our database, only throughput and latency measurements
- As a Nano-tier model, it is positioned below larger Nemotron 3 sizes and will generally trail them on the hardest reasoning tasks
- Reasoning models emit additional intermediate tokens, so billed output volume per request is typically higher than for a non-reasoning model answering the same prompt
- Provider availability for smaller Nemotron variants is often narrower than for widely hosted open models, which limits price and region shopping
Key Features
•Reasoning-tuned variant that emits explicit intermediate steps before a final answer
•Approximately 329.9 output tokens per second measured throughput (Artificial Analysis)
•Roughly 554 ms time to first token measured latency (Artificial Analysis)
•Nano size tier within NVIDIA's Nemotron 3 family
•30B total / ~3B active parameter naming convention indicating a sparse mixture-of-experts design
•"Omni" variant designation within the Nemotron 3 Nano line
•Candidate for self-hosted deployment given its small active parameter footprint
About Nemotron 3 Nano Omni 30B A3B Reasoning
Nemotron 3 Nano Omni 30B A3B Reasoning is a model released by NVIDIA as part of its Nemotron 3 family. The "Nano" designation places it at the small end of that family's size tiers, and the "30B A3B" naming follows the convention NVIDIA and other labs use for mixture-of-experts models with roughly 30 billion total parameters and about 3 billion active per token. The "Reasoning" suffix indicates the variant is tuned to produce explicit intermediate reasoning before a final answer, rather than responding directly.
The most concrete figures we track for this model come from Artificial Analysis throughput testing: approximately 329.9 output tokens per second with a time to first token around 554 ms. That combination of a small active parameter count and high measured generation speed is relevant for reasoning models specifically, since chain-of-thought output inflates the number of tokens generated per request — faster generation directly shortens wall-clock latency on multi-step problems. We do not currently track a confirmed context window, modality list, or benchmark accuracy scores for this variant, so check the provider's own documentation before relying on a particular input length or input type.
In practice, models in this size and speed class are typically deployed where reasoning quality matters but per-request cost and latency budgets are tight: structured data extraction with verification steps, agent loops that make many sequential calls, and self-hosted or on-prem deployments where a ~3B active-parameter footprint keeps serving cost low relative to dense models of similar total size. Compare it against other small reasoning variants on this page's pricing table, since provider choice affects both throughput and cost for the same weights.
Common Use Cases
This model fits workloads that need step-by-step reasoning at high request volume rather than maximum single-answer accuracy: agentic loops that chain many tool calls and benefit from fast generation, math and logic tasks where showing intermediate work improves correctness, structured extraction pipelines with a verification pass, and coding assistance on scoped problems. The measured ~330 tokens per second output rate and ~554 ms time to first token make it viable for interactive assistants where a reasoning model would otherwise feel slow, and the small active parameter count makes it a reasonable candidate for self-hosted or on-prem serving when data residency matters. For the hardest research-grade reasoning, long-document analysis, or tasks where you need a confirmed large context window, evaluate larger Nemotron 3 tiers or other reasoning models alongside this one.
Frequently Asked Questions
How much does Nemotron 3 Nano Omni 30B A3B Reasoning cost?
Cost varies by provider and by pricing model — some hosts bill separately for input and output tokens, and reasoning tokens count toward output volume. Because this is a reasoning variant, expect higher output token counts per request than a non-reasoning model would produce. See the pricing table on this page for current per-provider rates, and if you self-host, your cost is driven by GPU hourly rates and achieved throughput instead.
What is this model best used for?
Tasks that benefit from explicit multi-step reasoning but run at volume or under latency pressure: agent loops with repeated tool calls, math and logic problems, structured extraction with a verification step, and scoped coding tasks. Its measured throughput of roughly 330 output tokens per second helps offset the extra tokens a reasoning model generates.
What does "30B A3B" mean in the model name?
It follows a common naming convention for sparse mixture-of-experts models: roughly 30 billion total parameters with about 3 billion active per token. In practice that means memory requirements scale with the total parameter count while compute per token scales with the much smaller active count, which is why models named this way often serve faster and cheaper than dense models of similar total size.
How does it compare to non-reasoning models of similar size?
A reasoning-tuned variant spends extra output tokens working through a problem before answering, which usually helps on multi-step math, logic, and constraint-following tasks but increases both latency and billed output volume on simple prompts. If most of your traffic is short classification or formatting work, a non-reasoning model in the same size class is typically the better fit.
Do we know its context window and supported input types?
We do not currently track a confirmed context window or modality list for this variant. The "Omni" designation in the name suggests broader input handling, but you should confirm supported modalities and maximum input length in NVIDIA's or your serving provider's documentation before building around them.