Open-weight hybrid Mamba-Transformer MoE with ~3.5B active parameters for cost-efficient, high-throughput agentic reasoning and tool use.
Modality:
Text
PDF
model ID
nvidia/nemotron-3-nano-30b-a3b
Output Speed *
222.32
tok/s
Intelligence Index *
8.9
/ 100
Context Window *
262144
tokens
Input price
0.36
Anytoken
Output price
1.44
Anytoken
Nemotron 3 Nano 30B A3B: Sparse MoE Efficiency for High-Volume Agentic Reasoning
Nemotron 3 Nano 30B A3B is an open-weight large language model trained from scratch by NVIDIA, part of the Nemotron 3 family (Nano, Super, Ultra). It uses a hybrid Mamba-2 + Transformer Mixture-of-Experts architecture with 30B total parameters but only ~3.5B active per token, activating 6 of 128 experts. The design targets inference economics: NVIDIA reports up to 3.3x higher throughput than comparable open models. It fits teams building tool-calling agents, long-context pipelines, and math/reasoning workloads where cost-per-token and throughput are the governing constraints.
Integrate Nemotron 3 Nano 30B A3B via the AnyAPI.ai API
Performance
Where Tool-Augmented Reasoning and Throughput Converge
Nemotron 3 Nano is strongest as a tool-calling reasoning model running at high throughput. Giving it Python tools lifts AIME 2025 accuracy from 89.1% to 99.2%, NVIDIA's clearest evidence that external tools close most of its reasoning gap. Because only ~3.5B of 30B parameters activate per token, independent measurements show output speeds above 260 tokens/second on reasoning tasks. For production, this means agent pipelines that execute many tool-augmented reasoning steps can run at materially lower cost per completed task than dense models of similar size, provided tool integration is wired correctly.
Benchmarks
Nemotron 3 Nano Benchmarks: Strong Math and Tool Use, Modest Pure Coding
NVIDIA reports 89.1% on AIME 2025 without tools and 99.2% with tools, outscoring Qwen3-30B-A3B-Thinking (85.0%). On SWE-Bench (OpenHands) it reaches 38.8%, ahead of Qwen3-30B-A3B (22.0%) and GPT-OSS-20B (34.0%), though well below dedicated coding frontier models. Long-context fidelity holds up: NVIDIA cites RULER scores of 92.9% at 256K and 86.3% at 1M. On general knowledge (MMLU-Pro, ~78.3%) it trails Qwen3-30B-A3B's ~80.9%. Artificial Analysis measures ~266 tokens/second output and ~1.2s time to first token in reasoning mode.
Output Speed
*
222.32
tok/s
Intelligence Index
*
8.9
/ 100
MMLU *
Broad world knowledge and problem-solving
79
%
GPQA *
PhD-level scientific reasoning across physics, biology, chemistry.
76
%
HLE *
Adherence to multi-step structured instructions.
11
%
LiveCodeBench *
Tool-calling reliability in long agentic loops.
74
%
Technical Specifications
What the model supports
Nemotron 3 Nano is a text-in, text-out hybrid Mamba2-Transformer MoE model (30B total, ~3.5B active). It ships with a configurable reasoning mode toggled via a chat-template flag, so you can trade accuracy for lower latency. The most production-relevant distinction is context: NVIDIA's hosted model card lists more conservative input/output limits, while self-hosted configurations support up to 1M tokens (256K is the common default due to VRAM). Note that context window and maximum output are separate limits; verify your provider's exact caps before building long-output workloads.
Verified Specifications —
NVIDIA Nemotron 3 Nano 30B A3B (Reasoning)
*
Input modalities
Text
PDF
output modalities
Text
Context window
262144
tokens
Maximum output tokens
235929
Reasoning
Yes
Knowledge cutoff
December 2025
Pricing (standard)
0.36
AnyTokens in
/
1.44
AnyTokens out
Limitations & Trade-offs
1
Modest pure coding performance. On SWE-Bench (OpenHands) Nemotron 3 Nano scores 38.8%, which beats same-tier open models like Qwen3-30B-A3B (22.0%) but sits well below dedicated coding-focused frontier models. For autonomous software-engineering agents that must resolve real repository issues end to end, a specialist coding model will produce higher solve rates. Nemotron is a better fit when coding is one capability within a broader reasoning or tool-use workflow rather than the primary success metric.
2
Reasoning overhead and latency on long tasks. The model's strength comes from generating reasoning traces before answering, but this increases output tokens and end-to-end response time—independent measurements show tens of thousands of reasoning tokens and multi-second time-to-first-token on complex tasks. For latency-critical, short-answer interactions you should disable the reasoning flag, accepting a slight accuracy drop on harder prompts, or choose a non-reasoning model.
3
Context limits vary by deployment. The headline 1M-token context applies to supported self-hosted configurations; the default Hugging Face configuration is 256K due to VRAM, and NVIDIA's hosted model card lists more conservative input/output limits, with third-party APIs commonly offering 128K-262K. If your workload depends on true million-token context, you must verify and provision the specific configuration rather than assuming the maximum is available everywhere.
4
Text-only modality. Nemotron 3 Nano 30B A3B accepts and produces text only—it has no image, audio, or video input despite NVIDIA shipping a separate Omni variant with those modalities. Teams needing document OCR, chart understanding, speech, or vision in the same model should evaluate the Nemotron 3 Nano Omni variant or another multimodal model instead of assuming parity across the family.
Best-Fit Workloads
Where this model earns its place
01
Tool-calling agents
The clearest fit. Tool use lifts AIME accuracy from 89.1% to 99.2%, and the model posts competitive agentic scores (SWE-Bench 38.8%, BFCL v4 53.8%). Combined with high throughput from sparse activation, it suits multi-step agent loops that call calculators, code execution, retrieval, or external APIs. Production agents that run many tool-augmented steps per task benefit most, since cost-per-completed-task stays low.02
Math and quantitative reasoning
Nemotron 3 Nano shows a quantifiable lead on competition math, scoring 89.1% on AIME 2025 without tools and 99.2% with Python tooling, plus strong MiniF2F results. Applications involving step-by-step quantitative problem solving, formal reasoning, or data analysis pipelines benefit, especially when a code-execution tool is attached. Disable reasoning only for simpler prompts where latency outweighs accuracy.03
Long-context document and codebase analysis
With up to 1M-token context in supported configurations and NVIDIA-reported RULER scores of 92.9% at 256K and 86.3% at 1M, the model handles cross-document reasoning and large-codebase inspection in a single request. The Mamba-2 hybrid backbone keeps long-context inference efficient. Confirm your provider actually exposes the context length you need, since defaults are often lower than the 1M ceiling.04
High-volume, cost-sensitive inference
Because only ~3.5B of 30B parameters activate per token, Nemotron 3 Nano delivers up to 3.3x higher throughput than comparable open models and occupies a low-cost tier, making it economical for high-request-rate production traffic. It fits batch classification, large-scale summarization, and backend reasoning tasks where per-token economics dominate. For the lowest-latency single responses, pair it with reasoning disabled.Pricing in anytokens via AnyAPI
Input
0.36
₳
Output
1.44
₳
Cache write
—
₳
Cache read
—
₳
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04