NVIDIA officially announced two major releases on August 11, 2026 through its official blog, in a post authored by Kari Briski: the Nemotron 3.5 Lightning open model and the NeMo Switchyard routing library. As AI shifts from conversational chatbots to autonomous agents, the announcement directly answers a growing market demand for full control over where AI runs and how it is deployed and evolved. The release is significant because it positions NVIDIA firmly in the open model arena for agentic workloads, following Nemotron 3 Nano and reinforcing the company’s commitment to continually improving open models for greater accuracy, speed, and efficiency across edge devices, PCs, workstations, data centers, and the cloud.
The centerpiece of the release is Nemotron 3.5 Lightning, a fully customizable open model built for the high-volume, specialized tasks that power always-on agents inside larger multi-agent systems. It is a 30-billion-parameter mixture-of-experts model with only 3 billion active parameters, which makes it small enough to run efficiently while delivering frontier-level intelligence for tasks such as code review, tool use, security alert monitoring, and answering billing questions. According to NVIDIA, the model delivers up to 4x faster output speed, translating to roughly 30 percent faster agentic task completion compared with other models in its class, as demonstrated on PinchBench benchmarks. Alongside the model, NVIDIA introduced NeMo Switchyard, an open source model routing library that intelligently directs each request to the most capable and suitable model in a developer’s mix of open, proprietary, and NVIDIA models, automatically and without requiring applications to be rewritten.
Official details show the model was developed with contributions from the Nemotron Coalition, whose members provided evaluation methodologies, inference software, and datasets, and that NVIDIA published as much of the training data and techniques as licensing permits, alongside the Nemotron-RL-Agentic-Terminal-Pivot reinforcement learning dataset for coding agent capabilities. The model is available on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as an NVIDIA NIM microservice, through a broad ecosystem of cloud partners, and can run locally on RTX PCs, DGX Spark, DGX Station, and Jetson devices. Early adopters customizing the model include CrowdStrike for cybersecurity, Harvey with Trajectory for legal services, and CodeRabbit with Baseten for code review. NeMo Switchyard, available on GitHub and integrated by partners such as LangChain, Kong, Cognition, and Ramp, maintained frontier-level accuracy in internal benchmarks while reducing task completion cost to nearly one-third of running Opus 4.8 alone.
The impact of these releases extends across developers, enterprises, and the broader AI ecosystem. Teams building agentic applications gain the ability to pair a frontier reasoning model for planning with small specialized models for high-volume tasks, improving tokenomics while keeping data private on local hardware, and enterprises already report cost reductions of roughly 28 to 74 percent with maintained quality. For the industry, the release intensifies competition in open agentic models and signals that model routing, not just model size, is becoming a decisive layer of AI infrastructure, likely pushing rivals to offer comparable routing ecosystems. Going forward, NVIDIA is expected to expand the Nemotron 3.5 lineup, with reports suggesting an even larger Nemotron 4 model under development, while NeMo Switchyard rolls out to additional partner platforms, further entrenching NVIDIA’s influence across the agentic AI stack.
Source: Nvidia Blog