NVIDIA ships Nemotron 3.5 Lightning and NeMo Switchyard
NVIDIA has expanded its Nemotron model family with two releases targeting enterprise agentic AI workloads: Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts open model, and NeMo Switchyard, an open source library for intelligent prompt routing across heterogeneous model ensembles.
The announcements follow the earlier Nemotron 3 Nano release and reflect NVIDIA's broader push to supply the full model-and-orchestration stack for always-on autonomous agents, rather than competing solely on GPU hardware.
What is being released
Nemotron 3.5 Lightning is designed for high-volume, specialised tasks within larger multi-agent systems: code review, tool use, security monitoring and domain-specific question answering. NVIDIA says the model delivers up to four times faster output speed and around 30% faster agentic task-completion compared with other models in its parameter class, based on PinchBench benchmarks. Because the model is open and customisable, organisations can post-train it on proprietary data using NVIDIA NeMo without rebuilding their existing pipelines.
Named early adopters include CrowdStrike for cybersecurity workloads, Harvey and Trajectory for legal services, CodeRabbit and Baseten for code review, and Fastino Labs for software development, finance and healthcare tasks. NVIDIA is also releasing Nemotron-RL-Agentic-Terminal-Pivot, a reinforcement-learning dataset used to post-train the model for coding-agent capabilities.
NeMo Switchyard addresses a different problem: as enterprises build systems that mix frontier reasoning models with smaller specialist models, manually managing which model handles each request creates integration overhead. Switchyard routes each prompt automatically to the most capable and cost-efficient model available in a developer's stack, without requiring code changes to existing applications. NVIDIA internal benchmarks claim the library maintains frontier-level accuracy while reducing task-completion cost to approximately one-third of running Opus 4.8 alone. Partner results cited in the release include Ramp cutting costs by 58% and runtime by 33% on SWE-Bench, LangChain achieving 74% lower cost in multi-turn agent tasks, and Boomi reaching 100% domain-routing accuracy while diverting 59% of traffic to a five-times-faster fine-tuned model.
Market context and competitive positioning
The releases position NVIDIA more directly against the growing class of inference-optimisation vendors, including Anyscale, Together AI and Fireworks AI, as well as open-source routing projects gaining traction in the LangChain and LiteLLM ecosystems. Both of those communities are among the named NeMo Switchyard integration partners, which suggests NVIDIA is prioritising ecosystem compatibility over lock-in at the orchestration layer.
The mixture-of-experts architecture in Nemotron 3.5 Lightning mirrors the approach adopted by several frontier labs seeking to deliver high capability at lower inference cost. For enterprise buyers, the combination of an open, locally deployable model with a plug-in routing layer addresses two recurring procurement concerns: data-residency control and total cost of inference at scale.
Regulatory tailwinds add further urgency. The EU AI Act's general-purpose AI obligations, which began phasing in during 2025, incentivise enterprises to prefer open, auditable models over opaque proprietary systems. NVIDIA's policy of publishing training data and techniques "as far as licensing permits" is directly responsive to those traceability requirements, though the caveat leaves the scope of disclosure open to scrutiny.
Switchyard is available now as an open source library. Nemotron 3.5 Lightning is available for download, with support for NVIDIA RTX PCs, DGX Spark, DGX Station and Jetson hardware, as well as data-centre and cloud deployments. NVIDIA has not disclosed a standalone commercial support pricing model for either release.