Nvidia Releases Nemotron Lightning and NeMo Switchyard
Nvidia released Nemotron 3.5 Lightning, a 30-billion-parameter open mixture-of-experts model, and NeMo Switchyard, an open-source routing library for agentic AI workloads, on August 11. Nvidia describes Lightning as a specialized execution model for high-volume agent tasks, while Switchyard routes requests among models without requiring application rewrites. Nvidia reports up to 4x faster output versus comparable models in its class.
Nvidia released Nemotron 3.5 Lightning and NeMo Switchyard on August 11, pairing a 30-billion-parameter open mixture-of-experts model with an open-source library for routing AI-agent workloads across multiple models. The releases target long-running agent systems in which lower-cost specialized models handle frequent execution tasks while larger models handle more demanding reasoning.
According to Nvidia's developer documentation, Nemotron 3.5 Lightning has 3 billion active parameters and is designed for high-volume, low-latency tasks in always-on agent workflows. Nvidia lists tool use, result validation, code review, security-alert monitoring, billing questions and subagent delegation among the sorts of workloads that can be assigned to specialized models.
Nvidia reports that Lightning uses speculative decoding, harness-optimized training, and NVFP4 and BF16 checkpoints. The company claims up to 4x faster output than comparable models in its class and approximately 30% faster agentic task completion than Qwen3.6-35B at matched accuracy. Those performance figures are vendor-reported results, rather than independent benchmark findings.
Model routing accompanies the release
NeMo Switchyard is an open-source routing library intended to select a model for individual requests or steps within an agent workflow. Nvidia's announcement states that enterprises can build routers around their own requirements, and that Switchyard can direct requests to a suitable model without developers rewriting the surrounding application.
The technical premise is familiar in production agent systems: model quality, inference cost, latency, context needs and tool-use reliability vary by task. Nvidia's Switchyard documentation frames routing as a way to use specialized and frontier models together rather than treating model selection as a one-time architecture decision.
VentureBeat reports that Nvidia's own tests found a Switchyard-based combination could reduce benchmark costs to about one-third of running Anthropic's Opus 4.8 alone while retaining what Nvidia characterized as frontier-level task completion. The publication also identifies Not Diamond and the open-source RouteLLM project from UC Berkeley and LMSYS as relevant routing alternatives. These comparisons should be treated carefully because routing results depend heavily on the task mix, evaluation harness, candidate model set, pricing assumptions and failure tolerance.
Customization and deployment claims
Nvidia describes Lightning as customizable through post-training with NeMo using enterprise data, tools and workflows. SiliconANGLE reports that Kari Briski, Nvidia vice president of generative AI, cited a CodeRabbit example in which a router agent was trained for one epoch using Nvidia's standard auto-model recipe for $85 in about two hours. Briski also said another partner placed Lightning into an existing post-training stack with no changes required.
Those are examples provided by Nvidia and do not establish a general post-training cost or elapsed-time expectation. In comparable deployments, practitioners typically need to validate data-preparation work, evaluation coverage, serving hardware, quantization effects and model-routing fallback behavior before extrapolating from a narrowly scoped fine-tuning example.
The combined release is notable because Nvidia is shipping both a specialized open-weight model and routing software around it. For teams operating agentic systems, the relevant evaluation is not only a model benchmark. It is whether a router can preserve task completion and tool reliability across real workflow distributions while reducing expensive-model calls, operational complexity and latency.
Key Points
- 1Nvidia paired a 30B MoE execution model with an open routing library, addressing both model efficiency and multi-model orchestration.
- 2Nvidia reports 3B active parameters and up to 4x faster output, but teams should independently validate workload-specific throughput and accuracy.
- 3Comparable agent deployments increasingly evaluate routing policies, fallbacks, and tool reliability alongside model quality because workflow steps have different cost profiles.
Scoring Rationale
This is a notable release for teams building multi-model agent systems, combining an open specialized model with routing infrastructure. Its practical value depends on whether Nvidia's reported latency, cost, and task-completion results reproduce under enterprise workloads and evaluation harnesses.
Sources
Primary source and supporting public references used for this report.
View 3 more sources
- NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agentsdeveloper.nvidia.com
- Nvidia's Switchyard router reshuffles AI models mid-task, cutting task costs to a third in its own testsventurebeat.com
- Nvidia releases Nemotron 3.5 Lightning and NeMo Switchyard to give enterprise AI capability optionssiliconangle.com
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
