Nemotron Nano 2 · NVIDIA · Model History

Claims and evidence

Released 18 August 2025 Primary source[1]Published: August 18, 2025; Models [2]Release Date

Status AvailableNVIDIA has no deprecation page for its open models; the weights of all checkpoints were still downloadable from NVIDIA's Hugging Face organisation on 2026-10-01. Primary source[2] [1]Models

Successor of Llama-3.1-Nemotron-Nano-8B-v1 EditorialEditorial link along the Nano tier: Nemotron Nano 2 is the next general-purpose Nano release and carries the number 2; NVIDIA pages do not name a predecessor. Unlike the Llama-based v1 model, it was trained from scratch by NVIDIA. Primary source[1]

Change · Architecture Hybrid Mamba-2 and Transformer design trained from scratch by NVIDIA, whereas the v1 Nano model was derived from a Llama model.Compared with Llama-3.1-Nemotron-Nano-8B-v1 Primary source[2]
Change · Reasoning Adds a thinking budget that can be set at inference time, next to switching reasoning on or off.Compared with Llama-3.1-Nemotron-Nano-8B-v1 Primary source[1]
Change · Training data Most of the pre-training data was released alongside the model as the Nemotron Pretraining Dataset v1 collection.Compared with Llama-3.1-Nemotron-Nano-8B-v1 Primary source[1]

Input text Primary source[2]Input
Output text Primary source[2]Output
Feature Reasoning modeReasoning switched on or off with /think and /no_think; the thinking budget can be set at inference time. Primary source[2]Model Overview; Reasoning Budget Control [1]Technical Highlights, Post-training
Feature Function calling Primary source[2]Using Tool-Calling with a vLLM Server
Open weights YesReleased on Hugging Face under the NVIDIA Open Model License Agreement. Primary source[1]Models [2]License/Terms of Use; Release Date
Context window 128K tokens Primary source[1]Models [2]Input
Parameters 8.89BCount for the selected pruned architecture of the 9B models (Candidate 2, which the report says was used for Nano 2); the official model name rounds it to 9B. The 12B variants have their own figure. Primary source[3]Section 4.2, Table 10 (Candidate 2)
Access Open-weights download, APIHugging Face and the NVIDIA API Catalog (build.nvidia.com) on 08/18/2025. Primary source[2]Release Date

Not derived from another model.

None recorded.

Same dates as Nemotron Nano 2.

Variant Inline variant in this record.The aligned and pruned reasoning model, offered on Hugging Face and the NVIDIA API Catalog. Primary source[1]Models [2]Release Date

Also known as: Nemotron-Nano-9B-v2, NVIDIA-Nemotron-Nano-v2-9B

Same dates as Nemotron Nano 2.

Variant Inline variant in this record.Pruned base model, pruned and distilled from the 12B base model. Primary source[1]Models [4]Model Overview

Also known as: Nemotron-Nano-9B-v2-Base

Same dates as Nemotron Nano 2.

Variant Inline variant in this record.The base model before alignment or pruning, pre-trained on 20 trillion tokens in FP8. Primary source[1]Models [5]

Also known as: Nemotron-Nano-12B-v2-Base

Differs in:

Variant Inline variant in this record.Aligned 12B reasoning checkpoint published eleven days after the first three; NVIDIA builds Nemotron Nano 2 VL on it. Primary source[6]

Also known as: Nemotron-Nano-12B-v2

Differs in:

AI Radar coverage starts in June 2026; no coverage linked yet.