Nemotron 3 Nano is the pick for teams building agentic tool-calling pipelines rather than chasing the top reasoning score: giving it tools lifts AIME accuracy from 89.06% to 99.17%, and its MoE architecture keeps inference cheap at scale. It's the wrong choice if you need vision, audio, or frontier-level coding.
Nemotron 3 Nano 30B-A3B is NVIDIA's open-weight hybrid Mamba-Transformer Mixture-of-Experts model built for high-throughput agentic AI. It runs a native 1-million-token context window, far past the 262K default most hosted providers expose, and retains 68.2% accuracy on the RULER long-context benchmark at full length.
Where it sits
- $0.088/M$ per 1M tokensBlended price (3:1)Lower is better#5 / 64peer median $1.70/Mvendor price, checked by HokAI
- --tokens/sOutput speedHigher is better-- / 39peer median 90 tok/scited: Artificial Analysis
- 38.8%% solvedSWE-bench VerifiedHigher is better#28 / 28peer median 78.3%per source, see benchmark scores
- 73%% correctGPQA DiamondHigher is better#35 / 44peer median 88.3%per source, see benchmark scores
Cheaper than 92% of the 64 GA models with a published price, in the bottom third on SWE-bench Verified (rank 28 of 28), and one of 65 whose vendor states it does not train on customer data.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: NVIDIA · Family: Nemotron 3
Context window: 1,000,000 tokens
Input modalities: text, tool-calls · Output: text, tool-calls
About Nemotron 3 Nano 30B A3B
Nemotron 3 Nano 30B-A3B is an open-weight language model built by NVIDIA and released on December 14, 2025 as the entry tier of the Nemotron 3 family, ahead of the larger Super (120B, released March 2026) and Ultra (550B, released June 2026) tiers. It was designed to make agentic AI, tool-calling loops carrying long conversation histories, affordable at production scale rather than to chase the largest possible parameter count.
The model interleaves Mamba-2 state-space layers with grouped-query-attention Transformer layers inside a Mixture-of-Experts framework across 52 total layers, with each MoE layer routing to 6 of 128 experts per token plus shared experts active on every token. Of its 31.6 billion total parameters, only about 3.2 billion activate per token, roughly a tenth of the full weight count, the architectural lever behind its throughput. On other benchmarks from NVIDIA's technical report (arXiv 2512.20848), it scores 73.04% on GPQA, 68.25% on LiveCodeBench v6, and 71.51% on IFBench prompt-following.
Training used a 25-trillion-token pretraining corpus, including 2.5 trillion newly-added Common Crawl tokens, plus 13 million post-training samples across 15+ languages, with a Warmup-Stable-Decay learning-rate schedule; NVIDIA has not disclosed an exact training-data cutoff date. The model ships under the NVIDIA Open Model License, a commercially permissive custom license whose rights terminate automatically if a user disables the model's built-in safety guardrails.
It's available through NVIDIA's own build.nvidia.com hosted API, Hugging Face weight downloads, OpenRouter, and DeepInfra, or self-hosted via vLLM, SGLang, TensorRT-LLM, llama.cpp, LM Studio, or Unsloth. NVIDIA has signaled a coalition-built Nemotron 4 family as the next generation, with the Super and Ultra tiers already extending the same architecture through mid-2026.
Pricing
Input tokens run $0.05 per million and output tokens run $0.20 per million on both DeepInfra and OpenRouter; NVIDIA sets no official per-token price since this is an open-weight model. Self-hosting on an owned or rented NVIDIA H100 or H200 GPU is the lowest per-token cost at high sustained volume.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0015 | $0.0002 | $0.0017 |
| Support reply | $0.0001 | $0.0001 | $0.0002 |
| One coding agent run | $0.010 | $0.0040 | $0.014 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Hybrid Mamba-Transformer MoE architecture: 52 layers interleave Mamba-2 state-space blocks, GQA attention, and MoE routing across 128 experts (6 active per token plus shared experts), the structural reason for its throughput edge over dense models.
- Native 1-million-token context window: Extended through continued pretraining at a 512K sequence length; the RULER benchmark shows 68.2% accuracy retained at full context, ahead of Qwen3-30B-A3B at comparable depth.
- Configurable reasoning budget: A Reasoning ON/OFF toggle with an adjustable thinking-token budget lets developers trade accuracy for lower inference cost on a per-request basis.
- Native tool-calling and structured output: Tool-enabled prompting lifts AIME accuracy from 89.06% to 99.17%, NVIDIA's clearest demonstration that giving the model calculator or code-execution tools closes most of its reasoning gap.
- Open commercial license with released safety data: The NVIDIA Open Model License permits commercial redistribution of the model and its derivatives; NVIDIA also published roughly 11,000 labeled agent-safety traces from tool-using workflows.
Pros
- Built to run cheaply at scale rather than chase leaderboard position, exactly what high-volume agentic pipelines need.
- Commercially permissive open licensing removes the biggest practical barrier to self-hosting: no royalty or attribution headache for production deployments.
- The tool-enabled reasoning jump is a bigger swing than most models its size manage, and it matters more in practice than the base benchmark score.
Cons
- Text-only: it does not accept image, audio, or video, so multimodal work needs the separate Nemotron 3 Nano Omni variant instead.
- Modest 38.76% SWE-Bench (OpenHands) score versus dedicated coding-focused frontier models.
- Trails Qwen3-30B-A3B on general-knowledge benchmarks like MMLU-Pro, a real tradeoff for the throughput gain.
- Most managed API providers default to a 262K context window rather than the full native ceiling, so the larger window needs explicit configuration.
Benchmarks
- MMLU-Pro: 78.3% vendor-reported · 15 Dec 2025 — A harder version of the 57-subject knowledge exam, % correct.
- AIME 2025: 89.1% vendor-reported · 15 Dec 2025 — Competition-level maths problems from the 2025 exam, % solved.
- GPQA Diamond: 73.0% vendor-reported · 15 Dec 2025 — PhD-level science questions that are hard to search for, % correct.
- Livecodebench: 68.25 vendor-reported · 15 Dec 2025
- Ifbench Prompt: 71.51 vendor-reported · 15 Dec 2025
- SWE-bench Verified: 38.8% vendor-reported · 15 Dec 2025 — Real GitHub issues fixed end to end, % solved.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much do you pay for Nemotron 3 Nano 30B A3B?
It runs $0.05 per 1M input tokens and $0.20 per 1M output tokens on both DeepInfra and OpenRouter, among the cheapest rates for a model in this benchmark class, with no official NVIDIA-set price since it's open-weight. A 100K-token document summary costs roughly $0.007 at that rate, and a 1M-in/200K-out coding-agent session runs about $0.09. Self-hosting on a single NVIDIA H100 or H200 GPU removes per-token fees entirely and is the lowest-cost path at sustained high volume.
What do you get on Nemotron 3 Nano 30B A3B's free tier?
NVIDIA doesn't run a hosted free tier of its own, but OpenRouter offers rate-limited free access to the model, and the open weights themselves cost nothing to download and run under NVIDIA's open license. The only real expense in a self-hosted deployment is GPU time, since there's no royalty attached to commercial use.
What are Nemotron 3 Nano 30B A3B's closest competitors?
Qwen3-30B-A3B is the closest open-weight rival in the same 30B-class MoE tier and edges Nemotron out on general-knowledge benchmarks. GPT-OSS-20B is another open competitor NVIDIA benchmarks against, though NVIDIA's own report puts Nemotron's throughput 2.2x ahead of it on identical hardware. Neither rival is currently listed on HokAI; both are named directly in NVIDIA's Nemotron 3 technical report as the comparison set.
How does Nemotron 3 Nano 30B A3B compare to Qwen3-30B-A3B in 2026?
The two trade places depending on what you're optimizing for: Qwen3-30B-A3B scores higher on general knowledge (80.9% vs Nemotron's 78.3% on MMLU-Pro), while Nemotron claims 3.3x its throughput on identical single-H200 hardware at an 8K input / 16K output workload, credited to its hybrid Mamba-Transformer-MoE architecture activating fewer parameters per token. Both figures come from NVIDIA's own technical report rather than independent third-party benchmarking, so treat the throughput claim as a vendor number pending outside verification.
How long does it take to get going with Nemotron 3 Nano 30B A3B?
The fastest way in is an API key from OpenRouter, DeepInfra, or NVIDIA's own build.nvidia.com console, all of which host the model with no local setup required. To self-host instead, download the BF16, FP8, or 4-bit GGUF weights from Hugging Face and serve them with vLLM, SGLang, TensorRT-LLM, or llama.cpp, whichever your existing inference stack already runs. Either path gives you native tool-calling and the model's full context window out of the box, though most managed APIs default to a smaller window until you explicitly request the extended setting.