NVIDIA's Nemotron 3.5 Lightning Blends Mamba and MoE

NVIDIA has published Nemotron 3.5 Lightning 30B-A3B on Hugging Face, a text-generation and reasoning model that leans on an unusual architecture to keep inference costs down. It is a mixture-of-experts design with 30 billion total parameters but only about 3 billion active per token, released here in BF16 precision.

The standout detail is the hybrid layout: the model combines Mamba-style state-space components with a mixture-of-experts structure. That pairing is meant to deliver the throughput advantages of both approaches — Mamba's linear-scaling sequence handling and MoE's ability to route work to a small slice of the network at a time.

Built for agentic coding

NVIDIA positions the release specifically for agentic coding, the kind of multi-step, tool-using workflows where a model plans, edits, and iterates rather than answering a single prompt. A few things stand out for that use case:

• Only 3B parameters activate per token, which lowers the compute and memory bill for long, repeated inference loops.
• The reasoning modality suggests it is tuned to produce structured, multi-step outputs.
• The compact active footprint makes it a candidate for teams that want capable coding assistance without frontier-scale serving costs.

Why it matters: agentic tools call models constantly, so per-token efficiency compounds fast. A sparse, hybrid model that keeps quality up while cutting active compute is exactly the trade-off developers building autonomous coding agents have been looking for.

The model ships under a custom NVIDIA license rather than a standard open license, so teams should review the terms before deploying it. Context length and benchmark figures were not specified in the release record.