NVIDIA Nemotron Nano 2 and Nemotron-Pretraining v1

2025-08-18 · Publisher : NVIDIA · Domain : Artificial Intelligence

NVIDIA releases Nemotron Nano 2 (9B and 12B), a hybrid Mamba-Transformer reasoning model, along with most of its pretraining corpus (6.6 trillion tokens).

### Context After Nemotron-H and the Llama Nemotron models, NVIDIA wants a small, fast reasoning model that runs on a single entry-level GPU, and is starting to publish its pretraining data. The release comes as Qwen3-8B is the open reference of comparable size, against which NVIDIA measures itself. ### What the model brings - A 9B-parameter model obtained by pruning a 12B model pretrained on 20 trillion tokens in FP8. - A throughput claimed to be up to 6 times higher than Qwen3-8B, measured on a single A10G GPU in bfloat16, for comparable or better accuracy. - A 128K-token context that fits on an A10G (22 GiB of memory) according to NVIDIA. - A thinking budget adjustable at inference: about 5% of the post-training data contains truncated reasoning traces. - The release of Nemotron-CC-v2, Nemotron-CC-Math-v1 (133 billion tokens), Nemotron-Pretraining-Code-v1 (747.4 billion) and Nemotron-Pretraining-SFT-v1, for a total of 6.6 trillion tokens. ### Architecture and specifications Three models: NVIDIA-Nemotron-Nano-9B-v2 (aligned, pruned), 9B-v2-Base and 12B-v2-Base. Hybrid Mamba-2 and MLP architecture with four attention layers, 128K context, data cutoff September 2024. Languages (HF model card): English, German, Spanish, French, Italian, Japanese, Korean, Portuguese, Russian, Chinese, Arabic. License: NVIDIA Open Model License. Openness: weights and a large part of the pretraining corpus published, arXiv technical report 2508.14444; the full pretraining code was not verified, so this is a family with open weights and data, without the "fully open" qualification. ### Cost and access | Item | Detail | |---|---| | Official API | Not disclosed (trial on build.nvidia.com not verified for this model) | | Hardware | A10G GPU (22 GiB) in bfloat16, H100 80 GB, A100, Jetson AGX Thor (HF model card); FP8 (September 22, 2025) and NVFP4 (October 7, 2025) variants published afterwards | ### Benchmarks | Benchmark (reasoning on, published by NVIDIA) | Nano 9B v2 | Qwen3-8B | |---|---|---| | AIME25 | 72.1 | 69.3 | | MATH500 | 97.8 | 96.3 | | GPQA | 64.0 | 59.6 | | LiveCodeBench | 71.1 | 59.5 | | BFCL v3 | 66.9 | 66.3 | | IFEval (strict) | 90.3 | 89.4 | | RULER (128K) | 78.9 | 74.1 | Source: Hugging Face model card nvidia/NVIDIA-Nemotron-Nano-9B-v2, NVIDIA measurements. The comparison model is an open model. ### Versus closed models The published comparison is against an open model of similar size, not a closed model, and no independent ranking or comparison with a closed model was consulted. The only quantified gap is therefore NVIDIA's own against Qwen3-8B: a lead of 0.6 to 11.6 points depending on the test and a throughput up to 6 times higher. The model targets low-cost self-hosting, not parity with the large closed models. The availability of the corpus makes it possible to reproduce or adapt the pretraining, a sovereignty advantage that closed models do not offer. ### Key takeaways Nano 2 combines inference efficiency, an adjustable thinking budget and open pretraining data. It serves as the basis for the next generation, Nemotron 3.

Tags : LLM, Open Source, NVIDIA, Nemotron Nano, Mamba, Raisonnement, NVIDIA Open Model License