Nemotron 3 Nano 30B-A3B
A Mixture-of-Experts (MoE) architecture language model released by NVIDIA, with a total of approximately 30 billion parameters but only around 3 billion parameters activated per inference pass. With a storage size of approximately 24.57 GB, the model balances the knowledge capacity of large models with the inference speed of smaller ones.
Timeline (last 90 days)
The candidate lineup ranges from the 0.70GB LFM 1.2B to the 24.57GB Nemotron 3 Nano 30B-A3B.
In Nemotron 3 Nano 30B-A3B, 30B refers to a total parameter count of approximately 30 billion, while A3B indicates that approximately 3 billion parameters are actually activated per forward pass; the model weights still need to be fully loaded, and the runtime memory footprint is similar to that of a dense large model.
All Facts (2)
The candidate lineup ranges from the 0.70GB LFM 1.2B to the 24.57GB Nemotron 3 Nano 30B-A3B.
50%UnverifiedIn Nemotron 3 Nano 30B-A3B, 30B refers to a total parameter count of approximately 30 billion, while A3B indicates that approximately 3 billion parameters are actually activated per forward pass; the model weights still need to be fully loaded, and the runtime memory footprint is similar to that of a dense large model.
50%