NVIDIA launches open-weight model Nemotron 3 Nano 30B
NVIDIA has published details on Nemotron 3, a new open model family aimed at multi-agent and tool-using workflows. The first release is Nemotron 3 Nano, and NVIDIA is positioning it around high token throughput and a native 1M-token context window.
Nemotron 3 Nano is described as a hybrid Mamba-Transformer mixture-of-experts model. NVIDIA lists it as a 30B total and 3B active parameter model, while the Hugging Face write-up pegs it at 31.6B total with about 3.6B active per token. The same Hugging Face post says its router activates 6 of 128 experts per forward pass, and NVIDIA calls out targeting DGX Spark, H100, and B200 for deployment.
Source: NVIDIA
On speed, Hugging Face claims up to 4x faster inference than Nemotron Nano 2, and up to 3.3x higher throughput than other models in the same size class. It also publishes a single-GPU H200 example using an 8K input and 16K output setup, where Nano is listed at 3.3x the throughput of Qwen3-30B and 2.2x versus GPT-OSS-20B. NVIDIA also points to Artificial Analysis results, stating Nano scores 52 on the Intelligence Index v3.0 among similarly sized models.
Source: NVIDIA
The context window is the other headline metric. NVIDIA mentions a 1M-token context length , with the Hugging Face write-up describing a long-context extension stage that uses continued pretraining at 512K sequence length, mixed with shorter sequences to retain short-context benchmark behavior. NVIDIA frames the 1M window as support for large codebases, long documents, and long-running agent sessions without heavy chunking.
What’s new in Nemotron 3
- A hybrid Mamba-Transformer MoE backbone for superior test-time efficiency and long-range reasoning.
- Multi-environment reinforcement learning designed around real-world agentic tasks.
- A 1M-token context length supporting deep multi-document reasoning and long-running agent memory.
- An open, transparent training pipeline, including data, weights, and recipes.
- Immediate availability of Nemotron 3 Nano with ready-to-use cookbooks. Super and Ultra to follow.
Larger models in 2026
For deployment, NVIDIA lists cookbooks for vLLM, SGLang, and TensorRT-LLM, and also mentions running on GeForce RTX through tools like llama.cpp and LM Studio. The roadmap includes Nemotron 3 Super and Ultra in the first half of 2026 , with NVIDIA pointing to additions like latent MoE, multi-token prediction, and NVFP4 training. For context, NVIDIA’s older Nemotron-3 8B family (2023) was targeted for enterprise chat and Q&A use cases through NeMo and hosted catalogs.
Nemotron 3 Reinvents Multi-Agent AI With Efficiency and Accuracy
The Nemotron 3 family of MoE models includes three sizes:
- Nemotron 3 Nano, a small, 30-billion-parameter model that activates up to 3 billion parameters at a time for targeted, highly efficient tasks.
- Nemotron 3 Super, a high-accuracy reasoning model with approximately 100 billion parameters and up to 10 billion active per token, for multi-agent applications.
- Nemotron 3 Ultra, a large reasoning engine with about 500 billion parameters and up to 50 billion active per token, for complex AI applications.
Source: NVIDIA (Hugging Face)
