Nemotron 3 Nano 4B

Claims and evidence

• Released 16 March 2026Other sources give: 17 March 2026 [1]Published Publication date of the NVIDIA introduction article on Hugging Face. ● Primary source[2]Release Date

• Status AvailableNVIDIA has no deprecation page for its open models; the weights were still downloadable from the NVIDIA organisation on Hugging Face on 2026-10-01. ● Primary source[2]

• Derived from (distillation) Nemotron Nano 2Pruned from NVIDIA-Nemotron-Nano-9B-v2 with the Nemotron Elastic framework, then retrained by knowledge distillation from the frozen 9B parent. ● Primary source[1]Compressing 9B to 4B with Nemotron Elastic; Two-Stage Distillation for Accuracy Recovery [2]Model Overview; Model Architecture

• Change · Size Compressed from the 9B Nemotron Nano 2 model to about 4B parameters.Compared with Nemotron Nano 2 ● Primary source[1]Compressing 9B to 4B with Nemotron Elastic
• Change · Architecture Pruning cut depth from 56 to 42 layers and also reduced Mamba heads, feed-forward width and embedding size.Compared with Nemotron Nano 2 ● Primary source[1]Compressing 9B to 4B with Nemotron Elastic
• Change · Availability Built for on-device use on Jetson, GeForce RTX and DGX Spark; NVIDIA calls it its first model optimised for that purpose.Compared with Nemotron Nano 2 ● Primary source[1]
• Change · Training data Post-trained with a recipe derived from the Nemotron 3 post-training data, with reinforcement learning focused on instruction following and tool calling.Compared with Nemotron Nano 2 ● Primary source[1]

• Input text ● Primary source[2]Input
• Output text ● Primary source[2]Output
• Feature Reasoning modeOne model for reasoning and non-reasoning tasks; the reasoning trace can be turned off through the system prompt. ● Primary source[2]Model Overview
• Feature Tool use ● Primary source[1]
• Open weights YesReleased under the NVIDIA Nemotron Open Model License in BF16, FP8 and Q4_K_M GGUF. ● Primary source[2]License/Terms of Use [1]Try It Now!
• Context window 262K tokensStated as context length up to 262K. ● Primary source[2]Input
• Parameters 3.97 x 10^9The introduction article rounds this to 4 billion parameters. ● Primary source[2]Model Architecture
• Access Open-weights download, On devicePositioned for local deployment on Jetson Thor, Jetson Orin Nano, DGX Spark and RTX GPUs. ● Primary source[1]

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

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

All model releases from NVIDIA on AI Radar →