Nvidia Releases Nemotron 3.5 Lightning for Agents
The open 30B mixture-of-experts model activates 3B parameters per token and ships with training data, recipes and a new model router.
A smaller engine for long-running agents
Nvidia has released Nemotron 3.5 Lightning, an open-weight 30-billion-parameter mixture-of-experts model designed for high-volume and long-running agent workloads. Only about three billion parameters are active for each token, allowing the model to trade some maximum capability for lower latency and higher throughput.
The company says Lightning can deliver up to four times the throughput of comparable open models and reduce agent task-completion time by as much as 30%. Those are vendor measurements and will require independent testing across hardware and serving configurations. The model was available at launch through Hugging Face and multiple inference providers, while Ollama added support for local use. Nvidia released BF16 and optimized variants intended for different deployment and customization needs.
The release extends beyond weights. Nvidia is publishing the associated post-training data and recipes, including the Nemotron-RL-Agentic-Terminal-Pivot dataset used to teach terminal-oriented agent behavior. It also introduced NeMo Switchyard, an open-source routing library that can send individual steps in an agent workflow to different models. The underlying argument is that routine tool calls and intermediate decisions should not consume the same expensive model capacity as difficult reasoning steps.
A launch collaboration offered an early example of specialization. CodeRabbit and Baseten post-trained Lightning for code-review routing and reported that routing agreement rose from 75.8% to 80.4%. That narrow result does not establish general superiority, but it illustrates Nvidia’s intended workflow: start with an efficient open model, adapt it cheaply to a repetitive decision task, and deploy it close to the inference stack.
Why it matters
As agents run for minutes or hours and generate many intermediate tokens, economics increasingly favor model portfolios rather than sending every step to a frontier system. Nvidia is packaging the model, data, training workflow, router and optimized serving path as one stack. That strengthens its position above the GPU layer while giving developers a reproducible option for self-hosted agents. The central question is whether Lightning remains reliable on extended, failure-prone workflows; for agent deployments, a modest per-step error can compound faster than a throughput advantage.