NVIDIA announces the 30B Nemotron 3.5 Lightning model as open source

Summary

NVIDIA has released Nemotron 3.5 Lightning. This open Mixture-of-Experts (MoE) model features 30 billion total parameters but only 3 billion active inference parameters. Designed for continuously running AI agents, it claims to deliver large-model capabilities at small-model costs.

Why it matters

As agentic AI workloads shift to 24/7 operation, inference cost and latency have become critical constraints. This model’s stack of MoE, speculative decoding, and quantization serves as a direct engineering response to these challenges. Furthermore, publishing the weights, training data, and training recipes on Hugging Face under the OpenMDW-1.1 license lowers the barrier for teams looking to fine-tune or audit the model.

Highlights

• 3 billion active parameters out of 30 billion total via MoE routing
• Claimed output speed of ~670 tokens/sec, asserting leadership over competitors of similar size
• Fully open release: weights + training data + recipes under the OpenMDW-1.1 license

Critical take

The primary benchmark (24 points on Artificial Analysis, compared against OpenAI's gpt-oss-120b) comes from a single third-party index and a self-selected comparison. The 9-point increase over the previous Nemotron 3 Nano is modest. The article reads like a press release summary lacking independent replication or ablation studies. Consequently, the claim of 'small-model cost, big-model capability' remains unverified outside of NVIDIA’s own framework.