NVIDIA 发布 Nemotron 3.5 Lightning 与 NeMo

One day after Meta released Muse Glimmer, NVIDIA launched Nemotron 3.5 Lightning and NeMo Switchyard, showing a different approach to local agentic AI. Really cool.

Lightning is a 30B MoE with only 3B active parameters, distilled from Nemotron 3 Ultra and designed for the high-volume execution layer of always-on agents: tool calls, validation and subagent work.

NeMo Switchyard routes individual workflow steps across open and closed models. Complex planning can remain with a frontier model, while routine execution moves to Lightning. Its tuning-free routers can make those decisions without workload-specific training.

This is the same broad meaning of “agentic” as Meta’s Muse Glimmer: models operating tools across multi-step workflows.

The system design is different. Glimmer is a dense, multimodal local model built to understand screens, use tools and recover from failures. NVIDIA separates planning, execution and routing across a portfolio of models.

Meta is building the local agent brain. One could say: NVIDIA is building the local agent workforce. Open Source is so back :)