Nemotron 3.5 Lightning: 30B MoE for Agents

What exactly NVIDIA released

NVIDIA released not another near-frontier model, but a fast execution layer for agents: Nemotron 3.5 Lightning, an open 30B MoE model with 3B active parameters. In NVIDIA's official developer blog and on the NGC model card, it is described as a model for high-volume execution in long-running agents.

The meaning of MoE here is quite straightforward: total size 30B, but only 3B parameters are active per inference step. For agent systems this is more important than it sounds, because an agent often makes not one big request but hundreds of small calls: classify, extract a field, check state, perform a narrow step.

The NGC card names the architecture LatentMoE: a hybrid of Mamba-2, MoE, and Attention. This is an interesting mix specifically for execution, where you want not only quality on individual tasks but also decent throughput over long chains.

At announcement, NVIDIA claims up to 4× inference speed compared to models of similar size. I would keep this in mind as a vendor claim, not as an independent law of physics: such numbers always depend on hardware, batching, context length, and task type.

The most unusual part of the release is the openness of the package. According to NVIDIA materials, weights, data, and recipes are published under OpenMDW-1.1, and the NGC card separately states readiness for commercial use. For open agent infrastructure, this is stronger than just dumping weights and disappearing.

The main shift is not in intelligence, but in the execution loop

This release changes the game not for the top-level planner, but for cheap and frequent execution of subtasks. The NVIDIA hub explicitly describes a scenario where Lightning handles mass execution, while more powerful models stay for complex planning via NeMo Switchyard.

And here it got interesting: an agent system usually breaks not only on reasoning, but on the cost of each extra step. If the executor is slow, the agent starts looking smart only in demos, while in a real loop it turns into a token furnace.

Support through OpenClaw, Hermes Agent, and the NeMoClaw stack shows that NVIDIA is aiming not at single chats, but at the runtime layer. This is the right point of impact: a specialized executor should be boring, fast, and predictable.

What I would check first is not the shiny throughput, but stability on repetitive tasks with state errors. 4× speed helps little if the agent loops faster on a bad substep. Nemotron 3.5 Lightning looks like a practical brick for agent systems, but its main test will not be in a benchmark table, but in a long, dirty execution loop.

Previously we analyzed the performance graphs and context overhead of Claude Opus 4.6. Now, with the open MoE architecture of Nemotron 3.5 Lightning, these trade-offs become particularly clear — the model achieves a fourfold speedup while retaining 30 billion parameters.