Nemotron 3 Nano Explained
The 30B-A3B build activates about 3.5 billion parameters per token while handling up to 1M context.
Nemotron 3 Nano is the smallest tier of the open-weight Nemotron 3 family from NVIDIA. At Layer3Labs, we build and run AI systems inside other businesses, where selecting an open model depends on which build fits the graphics processing unit (GPU) you have. On NVIDIA's published benchmarks, its main model, Nemotron 3 Nano 30B-A3B, beats Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B on coding and instruction following, but trails Qwen3 on MMLU-Pro and Artificial Analysis Long Context Reasoning (AA-LCR).
Alongside the 30B model, the family provides Nemotron 3 Nano 4B for edge devices and Nemotron 3 Nano Omni for multimodal workloads.
What Nemotron 3 Nano Is
Nemotron 3 Nano serves as the entry-level tier of the Nemotron 3 family, sitting below the larger Super and Ultra editions. NVIDIA published the foundational model, Nemotron 3 Nano 30B-A3B, on Hugging Face on December 15, 2025. It uses a hybrid Mixture-of-Experts (MoE) architecture combining 23 Mamba-2 and MoE layers alongside 6 standard attention layers.
The 30B designation means the model contains 30 billion total parameters, but its MoE routing activates only about 3.5 billion parameters per token. Each MoE layer contains 128 experts and 1 shared expert, with 6 routed experts activated for any single token. Only the active parameters run for each token, so a token does not use all 30 billion parameters.
Parameter counts show a minor counting difference across official materials. The Hugging Face model card reports 3.5B active parameters and 30B total parameters. On the NVIDIA Nemotron 3 research page, NVIDIA documents 3.2B active parameters, or 3.6B when including embeddings, out of 31.6B total parameters. Both sets of figures represent NVIDIA's own internal measurements under different reporting conventions.
Nemotron 3 Nano operates as a single unified checkpoint for both reasoning and standard conversational tasks. It writes out an internal thinking trace before providing its final response by default. Developers who require direct generation can disable this behavior by setting enable_thinking to False in the chat template.
Training data for the 30B checkpoint includes a pre-training cutoff date of June 25, 2025, followed by post-training alignment data through November 28, 2025. The model natively supports English, German, Spanish, French, Italian, and Japanese, alongside standard programming languages.
Which Nemotron 3 Nano Should You Download?
Download the 30B-A3B for text agents and long context on one GPU, the 4B for edge devices and GeForce RTX machines, and Omni when you need video, audio or image input.
In our comparison of NVIDIA's three Nano model cards, two conditions stood out. The 30B-A3B supports 1M tokens but defaults to 256k on Hugging Face, and Omni uses a different license with a 256k limit.
| Build | Parameters | Inputs | Context | Released | License | Best Fit |
|---|---|---|---|---|---|---|
| Nano 30B-A3B (BF16, FP8, NVFP4 precisions) | 30B total, 3.5B active | Text | 1M tokens (256k default) | December 15, 2025 | NVIDIA Nemotron Open Model License | Text agents, coding, and long context on one enterprise GPU |
| Nano 4B (BF16, FP8, GGUF) | 3.97B total | Text | 262K tokens | March 16, 2026 | NVIDIA Nemotron Open Model License | Local assistants, edge devices, and desktop GeForce RTX cards |
| Nano Omni 30B-A3B (BF16, FP8, NVFP4) | 31B total, ~3B active | Video, audio, image, text | 256k tokens | April 28, 2026 | NVIDIA Open Model Agreement | Multimodal pipelines, optical character recognition (OCR), charts, and speech processing |
The standard Nemotron 3 Nano 30B-A3B fits text-centric agent execution and long-document retrieval. NVIDIA lists the H100-80GB and A100 as supported hardware. It supports up to 1 million tokens of context, though Hugging Face configurations default to 256k tokens to manage Video Random Access Memory (VRAM) overhead.
The Nemotron 3 Nano 4B model is a small language model with 3.97 billion parameters compressed from Nemotron Nano 9B v2, built as a hybrid of mainly Mamba-2 and multilayer perceptron (MLP) layers with four attention layers. It targets edge devices, gaming characters, local desktop assistants, and Internet of Things (IoT) hardware like Jetson Thor, DGX Spark, and GeForce RTX cards. It focuses on English and programming languages.
The Nemotron 3 Nano Omni model builds multimodal capabilities onto the 30B backbone. It accepts video clips up to 2 minutes in MP4 format, audio files up to 1 hour in WAV or MP3 formats, images, and text. Omni supports 256k context and operates under the NVIDIA Open Model Agreement rather than the standard Nemotron Open Model License.
How Good Is Nemotron 3 Nano?
On NVIDIA-reported benchmarks, Nemotron 3 Nano 30B-A3B leads Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B on SWE-Bench OpenHands, LiveCodeBench v6, IFBench, and RULER at 1M tokens, while trailing on MMLU-Pro, AA-LCR, and tool-free AIME25.
On the NVIDIA Nemotron 3 research page, NVIDIA reports that, with 8K input and 16K output tokens on a single H200 GPU, Nemotron 3 Nano delivers 3.3 times the inference throughput of Qwen3-30B-A3B and 2.2 times that of GPT-OSS-20B.
The following NVIDIA-reported evaluations compare Nemotron 3 Nano 30B-A3B against similarly sized open alternatives:
| Benchmark | Nemotron 3 Nano 30B-A3B | Qwen3-30B-A3B-Thinking-2507 | GPT-OSS-20B |
|---|---|---|---|
| MMLU-Pro | 78.3 | 80.9 | 75.0 |
| AIME25 (no tools) | 89.1 | 85.0 | 91.7 |
| GPQA (no tools) | 73.0 | 73.4 | 71.5 |
| LiveCodeBench v6 | 68.3 | 66.0 | 61.0 |
| SWE-Bench (OpenHands) | 38.8 | 22.0 | 34.0 |
| TauBench V2 average | 49.0 | 47.7 | 48.7 |
| IFBench (prompt) | 71.5 | 51.0 | 65.0 |
| AA-LCR | 35.9 | 59.0 | 34.0 |
| RULER-100 at 1M tokens | 86.3 | 77.5 | Not reported |
| RULER-100 at 256k tokens | 92.9 | 89.4 | Not reported |
On SWE-Bench OpenHands, Nemotron 3 Nano scores 38.8, leading Qwen3-30B-A3B at 22.0 and GPT-OSS-20B at 34.0. LiveCodeBench v6 shows a score of 68.3, surpassing both competitors.
Instruction following and long-context retrieval also show advantages. On IFBench prompt evaluations, Nemotron 3 Nano reaches 71.5 compared to 51.0 for Qwen3 and 65.0 for GPT-OSS-20B. On RULER-100 at 1 million tokens, it achieves 86.3 versus 77.5 for Qwen3.
However, Nemotron 3 Nano trails behind rival architectures in specific areas. Qwen3 scores higher on MMLU-Pro (80.9 vs 78.3) and leads substantially on the AA-LCR long-context reasoning benchmark (59.0 vs 35.9). On AIME25 without tools, GPT-OSS-20B leads at 91.7 compared to Nemotron 3 Nano at 89.1.
For the smaller 4B model, NVIDIA reports an AIME25 score of 78.5, MATH500 of 95.4, GPQA of 53.2, and LiveCodeBench of 51.8 with reasoning enabled. With reasoning turned off, it scores 82.8 on IFEval, 91.1 on RULER at 128k, and 61.1 on Berkeley Function Calling Benchmark (BFCL) v3.
These figures represent vendor-reported runs published by NVIDIA rather than independent third-party evaluations. You should test these checkpoints on your own prompt distributions before deploying them to production.
Running Nemotron 3 Nano Locally
The 30B model card documents Hugging Face Transformers, vLLM 0.12.0 or later, and SGLang, featuring a 256k default context that can be expanded to 1M in vLLM.
Supported serving engines include Hugging Face, vLLM, TensorRT-LLM, and SGLang. For the smaller 4B checkpoint, vLLM version 0.15.1 or later is required.
Context configuration requires attention to runtime settings. The default context length in Hugging Face configuration files is set to 256k tokens to mitigate excessive VRAM consumption. In vLLM, you can expand context to the full 1 million tokens by setting the environment variable VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 before startup.
Memory requirements vary across published precision checkpoints. NVIDIA distributes the 30B model in 16-bit brain floating point (BF16), 8-bit floating point (FP8), and NVIDIA 4-bit floating point (NVFP4) variants on Hugging Face. The 4B model provides BF16, FP8, and GGUF builds.
For multimodal serving, Nemotron 3 Nano Omni documents specific hardware floors. The BF16 Omni build has a 62 GB download size and requires a minimum of one NVIDIA H100 80GB GPU. The FP8 build is 33 GB and runs on one NVIDIA L40S 48GB GPU. The NVFP4 build is 21 GB and fits on a single NVIDIA GeForce RTX 5090 32GB GPU.
NVIDIA tested the 30B model across A100 80GB, H100 80GB, B200 192GB, RTX PRO 6000 96GB, Jetson Thor, and DGX Spark hardware. Developers planning self-hosted infrastructure can evaluate hardware capacity using our local AI hardware calculator. NVIDIA also offers a hosted trial on build.nvidia.com.
License and Commercial Use
Licensing terms differ depending on which Nemotron 3 Nano checkpoint your deployment relies upon.
Both the 30B-A3B and 4B models are released under the NVIDIA Nemotron Open Model License. The model cards explicitly state that these checkpoints are ready for commercial use.
In contrast, the Nemotron 3 Nano Omni multimodal model is distributed under the NVIDIA Open Model Agreement. This agreement is a distinct document with separate terms. Review each license document directly before integrating these checkpoints into commercial software or customer-facing infrastructure.
Who Should Not Use Nemotron 3 Nano
Teams needing maximum accuracy, organizations without a GPU and with light usage, and workloads requiring languages outside the six supported should not use Nemotron 3 Nano.
Teams that need maximum reasoning accuracy across complex enterprise workloads should consider larger tiers. Nemotron 3 Super provides 120 billion total parameters with 12 billion active parameters per token, requiring at least eight H100 80GB GPUs or 2 GPUs on B200 or B300 hardware. For maximum capacity, Nemotron 3 Ultra delivers 55 billion active parameters out of 550 billion total parameters.
Teams with no GPU and light usage may be better served by a hosted model than by running their own server. Furthermore, teams requiring native language support outside English, German, Spanish, French, Italian, and Japanese must look elsewhere, as those are the only human languages supported by the 30B model.
Our answer would change if independent benchmark results disagree with NVIDIA's table, if NVIDIA publishes a newer Nano release, or if the license terms change.
Nemotron 3 Nano is the right Nemotron 3 pick if you plan to self-host on a single GPU or need a 1M-token context on a budget. Choose the 4B build for small devices, Omni for video, audio, or image inputs, and Nemotron 3 Super or Ultra when accuracy matters more than hosting cost. That pick rests on NVIDIA's reported figures from its official model cards and research page.
To begin evaluating Nemotron 3 Nano for self-hosting, open the Nemotron 3 Nano 30B-A3B model card, select the BF16, FP8, or NVFP4 build that matches your available GPU memory, and test the model against your specific production workload.
Frequently Asked Questions
- Nemotron 3 Nano scores strongly on coding, instruction following, and long context according to NVIDIA-reported benchmarks. It achieves 68.3 on LiveCodeBench v6, 38.8 on SWE-Bench OpenHands, and 86.3 on RULER-100 at 1 million tokens, while trailing Qwen3 on MMLU-Pro and AA-LCR.
- Nemotron 3 Nano 4B is the smallest individual model, containing 3.97 billion parameters. At the family tier level, Nemotron 3 Nano represents the smallest overall classification, sitting below the Super and Ultra tiers.
- Nemotron 3 Nano demonstrates competitive coding performance for its parameter size. NVIDIA reports a 68.3 on LiveCodeBench v6 and 38.8 on SWE-Bench OpenHands, outperforming similarly sized models like Qwen3-30B-A3B and GPT-OSS-20B on those specific evaluations.
- The best Nemotron model depends on your hardware budget and workflow requirements. Nemotron 3 Nano is optimal for self-hosting on a single GPU or edge platform, whereas Nemotron 3 Super and Nemotron 3 Ultra provide superior reasoning capacity for teams with multi-GPU enterprise infrastructure.
- Nemotron 3 Nano Omni is a multimodal model built on the 30B backbone with roughly 3B active parameters per token. It processes text, images, audio up to 1 hour, and video up to 2 minutes, supporting a 256k token context window under the NVIDIA Open Model Agreement.
- Nemotron 3 Nano 30B-A3B supports up to 1 million tokens of input and output context. Hugging Face configurations default to 256k tokens to manage memory requirements, but developers can unlock the full 1 million tokens in vLLM using the VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 environment variable.
Planning an Open-Model Rollout?
Layer3Labs helps teams evaluate, deploy, and integrate open-weight models like Nemotron 3 Nano into automated business workflows with defined hardware guardrails.
Book a Consultation