NVIDIA-Nemotron-3-Nano-30B-A3B

Nemotron-3 Nano-30B is a new-generation LLM from NVIDIA. The model's key feature is its innovative architecture, which integrates Mamba2 layers, Transformer layers, and Mixture-of-Experts (MoE) technology into a unified compute cluster. This structure allows the model to efficiently process massive datasets while maintaining logical coherence and high throughput. The model has a total parameter count of 32 billion, but thanks to MoE routing, only an active subset of approximately 3.5 billion parameters is engaged for generating each individual token. This provides a unique balance: the model possesses the "knowledge" and capacity of a 30B-scale network but consumes computational resources on par with compact models optimized for fast inference. The model was trained on a dataset of about 25 trillion tokens, encompassing 43 programming languages and more than 19 natural languages.

Compared to Nemotron v2, the new version offers an MoE architecture instead of a dense one, delivering 4 times greater throughput. Another key capability of Nemotron-3 Nano is support for a context window of up to 1 million tokens. This expansion ideally showcases the capabilities of Mamba2 layers, which process long sequences with minimal memory overhead. A crucial stage in the model's creation was Multi-environment Reinforcement Learning using the NeMo Gym library. The model was trained not just to answer questions, but to perform action sequences: calling tools, writing functional code, and constructing multi-step plans. This makes its behavior more predictable and reliable in complex scenarios where step-by-step result verification is required.

On the AIME25 benchmark (American Invitational Mathematics Examination), which tests mathematical and quantitative reasoning, Nemotron 3 Nano achieves 99.2% accuracy with tool use, surpassing GPT-OSS-20B at 98.7%. On LiveCodeBench (v6 2025-08–2025–05), the model scores 68.2%, outperforming Qwen3-30B (66.0%) and GPT-OSS-20B (61.0%). On other benchmarks, the model either leads or is on par with its counterparts.

Given its architectural advantages and NVIDIA's recommendations, the model is ideally suited for the following tasks: Agentic Systems and Orchestration, Long-Context RAG, Local/On-Prem and Edge Computing, Code Generation, and Data Structuring.

Announce Date: 15.12.2025
Parameters: 32B
Experts: 128
Activated at inference: 4B
Context: 263K
Layers: 52, using full attention: 6, using no attention: 23
Attention Type: Hybrid Attention
Mamba Type: Mamba 2
Developer: NVIDIA
Transformers Version: 4.55.4
License: NVIDIA Nemotron Open Model License

Use our pre-built public endpoints for free to test inference and explore NVIDIA-Nemotron-3-Nano-30B-A3B capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.