NVIDIA's efficient hybrid model (30B total, 3.5B active MoE). Mamba-2 + Attention layers with 1M context for edge deployment.
Nemotron 3 Nano is NVIDIA's flagship open-weight small model in the Nemotron 3 family, released December 15, 2025. With 31.6B total parameters and only 3.6B active per token (including embeddings), it is a hybrid Mamba-2-Transformer Mixture-of-Experts (MoE) architecture optimised for agentic AI workloads — tool calling, structured outputs, multi-turn reasoning, and long-context processing. It supports a native 1M-token context window enabled by its efficient hybrid backbone, where Mamba-2 state-space layers handle long-range dependencies cheaply while sparse Transformer layers preserve relational precision.
For TheRouter operators, Nemotron 3 Nano is the most throughput-efficient open model in its parameter class. NVIDIA claims 3.3x higher throughput than Qwen3-30B-A3B and 2.2x higher than GPT-OSS-20B on a single H200 (8K/16K setting). It matches or exceeds comparable small MoE models on coding, math, and instruction-following benchmarks while costing only $0.06/M input tokens via TheRouter. The model comes with fully open weights (FP8 and BF16), training recipes, datasets (Nemotron-CC-v2.1), and the NeMo Gym RL environment suite — making it one of the most transparently documented models at its scale.
| Type | Rate |
|---|---|
| Input | $0.0648 / 1M tokens |
| Output | $0.2592 / 1M tokens |
| Release date | 2025-12-15research.nvidia.com ↗ | verified |
| Architecture | Hybrid Mamba-2-Transformer MoEresearch.nvidia.com ↗ | verified |
| Total parameters | 31.6Bresearch.nvidia.com ↗ | verified |
| Active parameters | 3.2B (3.6B with embeddings)research.nvidia.com ↗ | verified |
| Training data cutoff | June 25, 2025 (pre-training); November 28, 2025 (post-training)build.nvidia.com ↗ | verified |
| License | NVIDIA Open Model License (open weights, training recipe, and data released)huggingface.co ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MMLU-Pro Nemotron 3 Nano (Post-trained, CoT EM, 5-shot) — vs Qwen3-30B-A3B-Thinking-2507 at comparable settings | 65.05%% | research.nvidia.com ↗ | |
AGIEval-En Post-trained, CoT accuracy, 3/5-shot | 68.32%% | research.nvidia.com ↗ | |
HumanEval 0-shot pass@1 | 78.05%% | research.nvidia.com ↗ | |
MBPP-Sanitized 3-shot pass@1 | 75.49%% | research.nvidia.com ↗ | |
GSM8K 8-shot, accuracy | 92.34%% | research.nvidia.com ↗ | |
MATH 4-shot accuracy. Significantly ahead of comparable small MoE models (e.g. Qwen3-30B-A3B at 61.14%) | 82.88%% | research.nvidia.com ↗ | |
AIME 2025 Without tool assistance — surpasses Qwen3-30B-A3B (85.0%), slightly behind GPT-OSS-20B (91.7%) | 89.1%% | research.nvidia.com ↗ | |
GPQA Diamond Measured by Awesome Agents / public synthetic test (February 2026) | 84.2%% | awesomeagents.ai ↗ | |
RULER (64K) Long-context retrieval accuracy at 64K tokens — hybrid Mamba architecture advantage for long documents | 87.5%% | llm-stats.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "nvidia/nemotron-nano-30b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI-compatible chat completion — the simplest way to call Nemotron Nano via TheRouter.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia/nemotron-nano-30b",
"messages": [
{"role": "user", "content": "What are the key design goals of the Nemotron 3 architecture?"}
]
}'NVIDIA released the Nemotron 3 family of open models, led by Nano (31.6B total / 3.6B active) with a hybrid Mamba-2-Transformer MoE architecture, 1M-token context, and fully open weights, data, and training recipes.
NVIDIA published a detailed technical blog covering the hybrid MoE architecture, NeMo Gym multi-environment RL training, 1M-token context handling, and the transparent dataset pipelines that power Nemotron 3.
NVIDIA released the Omni variant of Nemotron-3-Nano, adding video comprehension, speech understanding, GUI interaction, and OCR capabilities while maintaining the compact 30B-A3B MoE footprint.
Built on Nemotron 3 Nano's architecture, Cascade 2 uses cascading agent pipelines to achieve GPT-OSS-120B-level results on coding and reasoning tasks while running entirely on a single consumer GPU.
Nemotron 3 Nano uses a hybrid Mamba-2-Transformer backbone instead of a pure Transformer. This lets it handle long contexts (up to 1M tokens) with far less memory overhead than dense attention, while the MoE layers keep the active parameter count at just 3.6B per token. The result is a model that rivals much larger models on benchmarks while being deployable on edge hardware.
On MATH (4-shot), Nemotron 3 Nano scores 82.88% vs 61.14% for Qwen3-30B-A3B. On AIME 2025, it reaches 89.1% vs 85.0% for Qwen3 and 91.7% for GPT-OSS-20B. NVIDIA claims 3.3x higher throughput than Qwen3-30B-A3B and 2.2x vs GPT-OSS-20B on a single H200. Overall, Nemotron is the strongest in math and throughput efficiency while GPT-OSS-20B leads slightly on competition-level math.
Yes. This was a primary design goal. The model was post-trained using RL across multiple interactive environments in NVIDIA's NeMo Gym — covering tool calling, multi-step planning, code execution, and structured output generation. TheRouter's supported_params include tools, tool_choice, and response_format, confirming full function calling capability.
Yes. NVIDIA released the model weights (FP8 and BF16), the base pre-trained checkpoint, the Nemotron-CC-v2.1 dataset (2.5 trillion tokens), Nemotron-CC-Code-v1 (428B code tokens), and the NeMo Gym RL environment suite under the NVIDIA Open Model License.
$0.06 per million input tokens and $0.16 per million output tokens. This is one of the lowest price points among supported models on TheRouter, reflecting the model's MoE efficiency (only 3.6B active parameters per token).
| source | URL | retrieved | |
|---|---|---|---|
| Release date | research.nvidia.com ↗ | 2026-05-28 | verified |
| Architecture | research.nvidia.com ↗ | 2026-05-28 | verified |
| Total parameters | research.nvidia.com ↗ | 2026-05-28 | verified |
| Active parameters | research.nvidia.com ↗ | 2026-05-28 | verified |
| Training data cutoff | build.nvidia.com ↗ | 2026-05-28 | verified |
| License | huggingface.co ↗ | 2026-05-28 | verified |
| MMLU-Pro | research.nvidia.com ↗ | 2026-05-28 | verified |
| AGIEval-En | research.nvidia.com ↗ | 2026-05-28 | verified |
| HumanEval | research.nvidia.com ↗ | 2026-05-28 | verified |
| MBPP-Sanitized | research.nvidia.com ↗ | 2026-05-28 | verified |
| GSM8K | research.nvidia.com ↗ | 2026-05-28 | verified |
| MATH | research.nvidia.com ↗ | 2026-05-28 | verified |
| AIME 2025 | research.nvidia.com ↗ | 2026-05-28 | verified |
| GPQA Diamond | awesomeagents.ai ↗ | 2026-05-28 | to verify |
| RULER (64K) | llm-stats.com ↗ | 2026-05-28 | to verify |
| NVIDIA Nemotron 3 family announced — Nano, Super, Ultra | research.nvidia.com ↗ | 2026-05-28 | verified |
| NVIDIA Technical Blog: Inside Nemotron 3 — Techniques, Tools, and Data | developer.nvidia.com ↗ | 2026-05-28 | verified |
| Nemotron-3-Nano-Omni released — multimodal variant with audio+video understanding | huggingface.co/nvidia ↗ | 2026-05-28 | verified |
| Nemotron-Cascade 2 — 30B open MoE runs one GPU, beats 120B models | awesomeagents.ai ↗ | 2026-05-28 | verified |
| What makes Nemotron 3 Nano different from other small MoE models? | developer.nvidia.com ↗ | 2026-05-28 | to verify |
| How does Nemotron 3 Nano compare to Qwen3-30B-A3B and GPT-OSS-20B? | research.nvidia.com ↗ | 2026-05-28 | to verify |
| Is Nemotron Nano good for agentic / tool-use applications? | developer.nvidia.com ↗ | 2026-05-28 | to verify |
| Is the Nemotron Nano model and its training data open source? | research.nvidia.com ↗ | 2026-05-28 | to verify |
| What pricing does TheRouter offer for Nemotron Nano? | www.therouter.ai ↗ | 2026-05-28 | to verify |