NVIDIA Nemotron Nano 2

What it is

Nemotron Nano 2 is a small reasoning model from NVIDIA, described in a paper NVIDIA posted on 20 August 2025^[1]. The version NVIDIA released for general use has 9 billion parameters^[1]. NVIDIA presents it as a unified model for reasoning and non-reasoning tasks: the same weights serve ordinary chat and deliberate problem solving, and in reasoning mode the model writes out a reasoning trace before it gives its final response^[3].

The point of a model this size is that it can be served cheaply. Much of the design follows from that goal.

A hybrid Mamba-Transformer architecture

The model is a hybrid of Mamba and Transformer layers, built on an earlier NVIDIA design called Nemotron-H. Its distinguishing choice is that most of the self-attention (self-attention) layers are replaced with Mamba-2 layers^[1].

Self-attention lets every token look at every other token in the sequence, so the computation and memory it needs grow quickly as the input and the output get longer. Mamba-2 layers belong to the state-space family: they carry a compressed running state along the sequence instead of comparing all pairs of tokens. Keeping a few attention layers and replacing the rest gives a model that can keep generating long outputs without the per-token cost rising as steeply. That matters especially for reasoning models, whose answers are long because the reasoning trace comes first.

How it was trained

The 9B model came out of a larger one. NVIDIA first pre-trained a 12B base model, Nemotron-Nano-12B-v2-Base, on 20 trillion tokens with an FP8 recipe^[1]. FP8 means carrying out the arithmetic in 8-bit floating point, a format that uses less memory and runs faster than higher-precision ones. That base model was then compressed, and its behaviour distilled into the 9B model, using the Minitron strategy^[1].

NVIDIA trained these models from scratch and lists Nemotron-CC-v2, Nemotron-CC-Math-v1 and the Nemotron post-training datasets among the data it used^[3]. The model card dates the model to June to August 2025 and puts the pretraining data cutoff at September 2024^[3].

Speed and context length

The efficiency claims are about inference, not training. In reasoning settings with 8K-token inputs and 16K-token outputs, NVIDIA reported up to 6 times the inference throughput of similar-sized models such as Qwen3-8B^[1]. Throughput is measured over the whole generated answer, so a model that reasons at length before answering benefits most from anything that makes each generated token cheaper.

The context window is the other half of the story. The model handles up to 128K tokens of context on a single NVIDIA A10G GPU with 22 GiB of memory^[1]. An A10G is a modest accelerator, so the claim is essentially that long-context inference does not require the largest hardware available.

What was released

NVIDIA published the 9B model together with the 12B base model and a 9B base checkpoint on Hugging Face, and released most of the pre-training and post-training datasets as well^[1]. Releasing the data alongside the weights lets others inspect what the model was trained on and fine-tune it for their own uses.

Where it sits

Nemotron Nano 2 was followed by Nemotron 3 Nano in December 2025, which NVIDIA said delivers up to 4 times the throughput of Nemotron 2 Nano^[2]. Seen as a step in that line, Nemotron Nano 2 is a model defined by trade-offs rather than raw scale: replace most attention layers with a cheaper sequence-mixing layer, compress a larger model into a smaller one, and publish the weights and data so the result can be checked and reused.