Nvidia ships Nemotron 3.5 Lightning and NeMo Switchyard
Nvidia is expanding its Nemotron 3 open model family with Nemotron 3.5 Lightning, a 30 billion parameter mixture of experts model built for specialized, high volume tasks inside larger multi agent systems, such as code review, tool use, security alert monitoring and answering billing questions. The company positions it as a worker model that sits underneath a larger frontier model, such as Nemotron 3 Ultra or GPT-5.6, which handles planning and orchestration while Lightning executes the repetitive steps. Nvidia says Lightning delivers up to 4x faster output speed than other models in its class, leading to 30% faster agentic task completion; the identity of those comparison models is not stated. It was developed with contributions from the Nemotron Coalition, whose members supplied evaluation methodologies, inference software and datasets, and it can be post trained with Nvidia NeMo on an organization's own data, tools and workflows. Nvidia is also releasing Nemotron-RL-Agentic-Terminal-Pivot, an agentic reinforcement learning dataset used to post train Lightning for coding agent capabilities, and, as with prior Nemotron releases, is publishing as much of the training data and technique as licensing allows. Companies customizing Lightning for their own workloads include CrowdStrike for cybersecurity, Harvey with Trajectory for legal services, and CodeRabbit with Baseten for code review; Lila Sciences is using it to improve reasoning for agentic tasks in physical and life sciences, and Fastino Labs customized it and reports leading accuracies in software development, finance and healthcare workloads. Lightning can run locally on NVIDIA RTX PCs, DGX Spark, DGX Station and Jetson devices, or scale to RTX PRO workstations, data centers and the cloud, and is available on Hugging Face, ModelScope, OpenRouter and build.nvidia.com as an NVIDIA NIM microservice, plus through Nvidia's cloud partner and inference platform ecosystem. Alongside it, Nvidia released NeMo Switchyard, an open source model routing library for agents. Rather than a developer picking one default model, which Nvidia argues either overspends or loses quality, or hand building routing logic, which is integration work that slows deployment, Switchyard automatically directs each step of an agent workflow to whichever model, open, proprietary or Nvidia's own, best fits the required quality, latency and cost. Nvidia's internal benchmarks show Switchyard holding frontier level accuracy while cutting task completion cost to nearly one third of using Opus 4.8 alone. Nvidia lists results from partners testing Switchyard: Boomi evaluated it across five routing capabilities and reported 100% domain routing accuracy, sent 59% of traffic to a 5x faster fine tuned model, and cut later turn latency by 21%. Cadence improved efficiency by 9.9% using the ChipStack AI Super Agent for formal verification. Classmethod's initial testing of opencode and Fireworks workloads through Switchyard showed a 27% cost reduction while maintaining quality. Cognition integrated Switchyard's staged router into Devin Desktop for Nvidia's internal use and cut mean cost by 28% relative to routing every request to a single frontier model, while staying near frontier performance on FrontierCode Main. Kong now delivers routing with Switchyard natively through its AI Gateway. LangChain achieved 74% lower cost across 145 multi turn Deep Agents tasks by routing only 7% of calls to a frontier model, at a 6% accuracy tradeoff. LiteLLM is adding Switchyard as a plug in to its proxy layer so developers get the benefit without changing their existing stack. Nous Research integrated it into Hermes. Ramp matched a frontier model's performance while cutting cost by 58% and runtime by 33% on Ramp SWE-Bench. Siemens is benchmarking it to improve efficiency in its Fuse EDA AI Agent. NeMo Switchyard is available now on GitHub, with support for additional partner platforms described as coming soon.
Key facts
- Nemotron 3.5 Lightning is a 30 billion parameter mixture of experts model that Nvidia says is up to 4x faster in output and 30% faster on agentic task completion than other models in its class
- NeMo Switchyard, Nvidia's new open source routing library, cuts task completion cost to nearly one third of using Opus 4.8 alone in Nvidia's internal benchmarks while keeping frontier level accuracy
- Partners report concrete gains from Switchyard: Boomi hit 100% domain routing accuracy while sending 59% of traffic to a 5x faster model, LangChain cut cost 74% by routing only 7% of calls to a frontier model, and Ramp cut cost 58% and runtime 33% on Ramp SWE-Bench
- CrowdStrike, Harvey with Trajectory, and CodeRabbit with Baseten are customizing Nemotron 3.5 Lightning for cybersecurity, legal work and code review respectively
- Nemotron 3.5 Lightning is live on Hugging Face, ModelScope, OpenRouter and build.nvidia.com as an NVIDIA NIM microservice; NeMo Switchyard is on GitHub now, with partner platform support coming
Why it matters
Agentic AI is moving from single chatbot calls to always on systems of multiple specialized models working together. Nvidia's release addresses two separate needs created by that shift: a small, fast, open model built specifically for the high volume repetitive subtasks inside an agent workflow, rather than a general purpose model handling everything, and a routing layer that picks the right model for each step automatically instead of a developer hard coding it. Used together, a frontier model such as Nemotron 3 Ultra or GPT-5.6 can plan and orchestrate a task while Lightning, or another model chosen by Switchyard, executes the narrow steps such as code review, tool calls, security alert triage or billing questions.
Who it affects
Companies building or operating agent systems at scale. Nvidia names CrowdStrike (cybersecurity), Harvey with Trajectory (legal services), CodeRabbit with Baseten (code review), Lila Sciences (reasoning for physical and life sciences agents) and Fastino Labs (software development, finance and healthcare) as customizers of Nemotron 3.5 Lightning. For NeMo Switchyard, named partners are Boomi, Cadence, Classmethod, Cognition (Devin Desktop), Kong (AI Gateway), LangChain, LiteLLM, Nous Research (Hermes), Ramp and Siemens (Fuse EDA AI Agent).
How to use it
Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, OpenRouter and build.nvidia.com as an NVIDIA NIM microservice, and through Nvidia's cloud partner, post training and inference platform ecosystem. It can be post trained with NVIDIA NeMo on an organization's own domain data, tools and workflows, and runs locally on NVIDIA RTX PCs, DGX Spark, DGX Station and Jetson devices, or scales to RTX PRO workstations, data centers and the cloud. NeMo Switchyard is open source on GitHub today; developers can tune or modify the router to prioritize quality, latency or cost, and LiteLLM is adding it as a plug in to its existing proxy layer so teams do not need to change their stack. No pricing or licensing terms are given for either release, and no date is stated for when Switchyard reaches additional partner platforms beyond GitHub.
How solid is it
The performance numbers come from Nvidia's own release post: its internal benchmarks for both the Lightning speed claims and the Switchyard cost claim against Opus 4.8, plus figures partner companies reported to Nvidia for inclusion in the post (Boomi, Cadence, Classmethod, Cognition, LangChain, Ramp). None of these are independent third party benchmarks, and they were measured under different tasks and conditions, Boomi's five routing capabilities, LangChain's own 145-task Deep Agents set, Ramp's SWE-Bench run, so the percentages are not directly comparable across partners. Nvidia does not name the models it compared Lightning against for the 4x and 30% figures, and gives no standalone accuracy benchmark score for Lightning itself, only relative speed and cost comparisons.
Risks and caveats
No pricing or licensing cost is disclosed for either Nemotron 3.5 Lightning or NeMo Switchyard. The article gives no calendar date for the announcement, only 'Today', and no individual at Nvidia or any partner is quoted or credited by name. The baseline models used for Lightning's speed comparisons are unnamed, which makes the 4x and 30% figures hard to verify independently. Because every efficiency and cost figure in the piece comes from Nvidia or from partners describing their own deployments, the numbers should be read as vendor and early adopter reporting rather than audited, apples to apples benchmarks.