Nemotron 3 Nano 4B
NVIDIA's 4B edge model, a Mamba-Transformer hybrid compressed from Nano 9B v2, with switchable reasoning for RTX PCs and Jetson.
In short. Nemotron 3 Nano 4B is a small AI model from NVIDIA meant to run on a gaming PC, a Jetson board or a DGX Spark. It can answer directly or think step by step, handles tool calls, and fits in about 8 GB. The license allows commercial use.
Summary
NVIDIA released Nemotron 3 Nano 4B on 16 March 2026. It has about 3.97 billion parameters and was compressed from NVIDIA-Nemotron-Nano-9B-v2 with the Nemotron Elastic framework. The architecture is a Mamba-2 and MLP hybrid with only four attention layers, which keeps memory use low at long context. The default configuration allows 262,144 tokens. The BF16 checkpoint is about 8 GB, an FP8 version is about 5.3 GB, and a GGUF build of about 2.8 GB is published for llama.cpp-style runtimes.
The model card positions it for agents on edge devices: gaming companions, local voice assistants and IoT automation, on Jetson Thor, GeForce RTX and DGX Spark. It is a single model that either reasons first or answers directly, controlled with a flag in the chat template, and it supports English and coding languages. NVIDIA recommends temperature 1.0 and top-p 0.95 for reasoning and temperature 0.6 for tool calling.
The card's own results include IFEval 82.8 prompt-level with reasoning off, and AIME25 78.5, MATH500 95.4, GPQA 53.2 and LiveCodeBench 51.8 with reasoning on. They are NVIDIA's numbers, run with NeMo-Skills. It uses the NVIDIA Nemotron Open Model License, which allows commercial use. For a newer small NVIDIA model, look at Nemotron 3.5 Bolt E4B.
Good for
- On-device agents and assistants on RTX PCs, Jetson and DGX Spark
- Game NPCs and local voice assistants that need fast answers
- Tool calling at small size, with reasoning switched on only when needed
- Fine-tuning a small hybrid model for a narrow task
Watch out for
- English and code only; other languages are not listed as supported
- It needs trust_remote_code and, for vLLM, version 0.15 or newer
- It is a small model, so it is weaker than the 30B and larger Nemotron models on hard reasoning
- The benchmark numbers come from NVIDIA
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Nano 4B (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Nano 4B (BF16) BF16 | 3.97B | 9.13 GB | 4.57 GB | 2.74 GB | Laptop or phone |
| Nano 4B (FP8) FP8 | 3.97B | 9.13 GB | 4.57 GB | 2.74 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
A permissive, Apache-style license from NVIDIA. Commercial use, derivative works and redistribution are allowed, NVIDIA claims no ownership of outputs, and the main duties are to pass the license on and keep notices.
See also
Frequently asked questions
- What is Nemotron 3 Nano 4B?
- Nemotron 3 Nano 4B is a small AI model from NVIDIA meant to run on a gaming PC, a Jetson board or a DGX Spark. It can answer directly or think step by step, handles tool calls, and fits in about 8 GB. The license allows commercial use.
- Can I use Nemotron 3 Nano 4B commercially?
- Yes. Nemotron 3 Nano 4B is released under NVIDIA Nemotron, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Nemotron 3 Nano 4B open source?
- Nemotron 3 Nano 4B is open-weight, not open source in the strict sense. You can download the model files, but the NVIDIA Nemotron adds conditions that the Open Source Initiative definition does not allow.
- Can I run Nemotron 3 Nano 4B on my own computer?
- Yes. The smallest variant, Nano 4B (BF16), needs about 2.74 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Nemotron 3 Nano 4B released?
- Nemotron 3 Nano 4B was released by NVIDIA on March 16, 2026.
- How long a context window does Nemotron 3 Nano 4B have?
- Nemotron 3 Nano 4B supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Nemotron 3 Nano 4B BF16 model card (Hugging Face) huggingface.co
- Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs arxiv.org
- NVIDIA Nemotron 3 family of models (NVIDIA Research) research.nvidia.com
- NVIDIA Nemotron Open Model License nvidia.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Nemotron 3 NVIDIA | 31.6B – 120B 12B active | 1M | NVIDIA Nemotron (Permissive) | 21 GB Gaming GPU | |
| Nemotron 3 Nano Omni NVIDIA | 31B 3B active | 256K | NVIDIA Open Model (Permissive) | 32 GB Gaming GPU | |
| Nemotron 3.5 NVIDIA | 8B – 120B 12B active | 1M | OpenMDW-1.1 (Permissive) | 5.52 GB Laptop or phone | |
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone |