NVIDIA Nemotron 3 (Nano 30B-A3B)

2025-12-15 · Publisher : NVIDIA · Domain : Artificial Intelligence

NVIDIA introduces the Nemotron 3 family (Nano, Super, Ultra) and releases Nano, a hybrid MoE with 30 billion parameters, 3 billion of them active, together with weights, datasets and training recipes.

### Context After Nano 2, NVIDIA moves to a Mamba-Transformer architecture combined with a mixture of experts, and announces three sizes: Nano (30B), Super (about 100B) and Ultra (about 500B). Only Nano is available at launch; Super and Ultra are announced for the first half of 2026. The stated focus is multi-agent agentic AI, with a lower cost per token. ### What the model brings - A throughput announced as 4 times higher than Nemotron 2 Nano and up to 60% fewer reasoning tokens, according to NVIDIA. - A native context of 1 million tokens (256K by default in the Hugging Face configuration). - A hybrid MoE: 23 Mamba-2 and MoE layers plus 6 attention layers, with 128 routed experts and 1 shared expert, 6 experts active per token. - Multi-environment reinforcement learning with the NeMo Gym library, released with NeMo RL and NeMo Evaluator. - Open datasets (the Nemotron pre-training and Nemotron post-training v3 collections), with 13 million post-training samples according to the technical blog. ### Architecture and specifications 30 billion parameters in total, about 3 billion active (3.5 billion according to the HF model card). Pre-training on about 25 trillion tokens, pre-training data cutoff June 25, 2025, post-training November 28, 2025. Languages: English, German, Spanish, French, Italian, Japanese (more than 20 languages and 43 programming languages seen during training). License: NVIDIA Nemotron Open Model License (commercial use allowed under conditions). Openness: weights, part of the pre-training data, post-training data, RL environments and recipes; NVIDIA states 3 trillion tokens of datasets published out of a pre-training corpus of about 25 trillion, so the data is partial. Nano is also available on AWS Bedrock and from several hosting providers (Baseten, DeepInfra, Fireworks and others). ### Cost and access | Item | Detail | |---|---| | Official API | NIM and build.nvidia.com; price per million tokens not disclosed at launch | | Hardware | A100 80 GB, H100 80 GB, B200 192 GB, RTX PRO 6000 96 GB, Jetson Thor, DGX Spark (HF model card) | | Weight memory | About 63 GB in BF16 (31.6 billion safetensors parameters, calculated), FP8 (December 6) and NVFP4 (December 20) variants | ### Benchmarks | Nemotron 3 Nano benchmark (published by NVIDIA) | Score | |---|---| | MMLU-Pro | 78.3 | | AIME25 (with tools) | 99.2 | | LiveCodeBench | 68.3 | | Arena-Hard-V2 (average) | 67.7 | | RULER-100@1M | 86.3 | Source: Hugging Face model card nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16, NVIDIA measurements, with no comparison column in the excerpt consulted. ### Versus closed models No comparison table with closed models was consulted for this model: the model card gives absolute scores and the NVIDIA press release is limited to throughput comparisons with its previous generation (4 times) and an efficiency ranking by Artificial Analysis, the details of which were not recorded. The positioning is that of a small, fast model for agents, not that of a direct competitor to the large closed models. The advantages put forward are transparency (data, recipes) and self-hosting on a single GPU. A score of 99.2 on AIME25 is obtained with tools, which is not comparable to the scores without tools in other announcements. ### Key takeaways Nemotron 3 Nano marks NVIDIA's shift to an open model whose entire chain (data, RL environments, libraries) is published. It announces Super and Ultra, delivered in March and June 2026.

Tags : LLM, Open Source, MoE, NVIDIA, Nemotron 3, Poids, données et recettes ouverts, NVIDIA Nemotron Open Model License