NVIDIA expanded its Nemotron 3 family on 2026-08-11 with Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts open model, and NeMo Switchyard, an open-source routing library for agent workflows. Nemotron 3.5 Lightning went up on Hugging Face, ModelScope, OpenRouter and build.nvidia.com, and NVIDIA also published Nemotron-RL-Agentic-Terminal-Pivot, the agentic reinforcement-learning dataset used to post-train it for coding-agent work. As with previous Nemotron launches, NVIDIA says it releases as much training data and method detail as licensing permits.
The framing is explicitly architectural rather than about a single best model. NVIDIA argues that always-on agents increasingly operate as “systems of models,” with a frontier reasoner such as Nemotron 3 Ultra or GPT-5.6 planning a workflow while smaller specialists handle targeted steps like code review, tool use, security-alert monitoring or billing questions. Lightning is positioned as that specialist. The company says it delivers up to 4x faster output speed and 30 percent faster agentic task completion than other models in its class, and that it can be post-trained on an organization’s own data with NeMo. It runs on RTX PCs, DGX Spark, DGX Station and Jetson as well as workstations, data centers and cloud - the on-premises and edge path is the point.
NeMo Switchyard is the other half. It routes each step of an agent workflow to whichever model - open, proprietary or NVIDIA’s own - best fits the developer’s stated priority of quality, latency or cost, without rewriting the application. NVIDIA’s internal benchmarks claim Switchyard “maintains frontier-level accuracy while reducing task completion cost to nearly one-third of Opus 4.8 alone.” Partner figures NVIDIA published are more informative because they are specific and self-reported by third parties: LangChain says it cut cost 74 percent across 145 multi-turn Deep Agents tasks by sending only 7 percent of calls to a frontier model, at a 6 percent accuracy tradeoff; Cognition reports 28 percent lower mean cost in Devin Desktop; Ramp reports 58 percent lower cost and 33 percent shorter runtime on its own SWE-Bench variant; Classmethod reports 27 percent. Kong, LiteLLM and Nous Research are integrating it into their gateways and agents.
The strategic read is that NVIDIA is standardizing the layer above the model, and doing it in open source. If routing becomes a library rather than a product, the value of any single frontier model in an agent pipeline falls, because the pipeline can substitute a cheap specialist for most steps - which is exactly the LangChain number, and exactly the outcome that sells more inference across more device classes. The accuracy tradeoffs are real and the headline efficiency claims are vendor or partner numbers with no independent replication, so treat the specific percentages as directional. The durable point for a technical leader is that “which model” is becoming a per-step routing decision rather than a procurement decision.