It’s open season for open-source artificial intelligence (AI).
NVIDIA Corp. joined the fray Tuesday with two major additions to its AI software lineup: Nemotron 3.5 Lightning, a lightweight open-source model designed for autonomous agents, and NeMo Switchyard, an intelligent model-routing tool engineered to slash enterprise operational costs.
The move lands as a fierce debate over open-source AI unfolds across Silicon Valley and Washington.
Following NVIDIA CEO Jensen Huang’s recent public defense of open-weight models, the chipmaker’s latest release signals a calculated strategy: driving down the cost of software implementation to ultimately fuel global demand for its flagship hardware.
Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model built for high-volume, specialized tasks within multi-agent systems. Distributed freely for download, adaptation, and commercial modification without licensing fees, the model is accessible via Hugging Face, ModelScope, OpenRouter, and NVIDIA’s platform.
Engineered specifically to run locally on hardware as compact as a single PC GPU, RTX systems, or scaled across enterprise data centers, Nemotron 3.5 Lightning delivers up to four times the output speed and 30% faster task completion compared to competitors in its weight class.
To achieve this performance, NVIDIA utilized distillation, a technique that trains smaller models using the outputs of larger, frontier-level systems. Kari Briski, NVIDIA’s vice president of generative AI, noted that early partners praise the model’s low barrier to entry. AI code-review platform CodeRabbit reportedly fine-tuned a custom router agent using Nemotron 3.5 Lightning in just two hours for $85 on a single graphics card.
Alongside the weights, NVIDIA is releasing its post-training datasets and recipes, allowing enterprises to blend proprietary data with NVIDIA’s foundational assets. Early enterprise testers include CrowdStrike Holdings Inc., CodeRabbit, and legal-tech developer Harvey.
Addressing the growing complexity of enterprise AI, NVIDIA also introduced NeMo Switchyard, an open-source model-routing library available on GitHub.
As workflows shift toward autonomous agents that perform mixed sequences — from simple sorting to multi-step reasoning — NeMo Switchyard automatically evaluates prompt requirements and directs tasks to the most cost-efficient and capable model.
Developers can customize routing criteria based on latency, budget, or accuracy metrics. Internal benchmarks demonstrate that NeMo Switchyard can maintain frontier-level precision while cutting overall task completion costs to roughly one-third of using Anthropic’s Opus 4.8 alone. Early adopters report substantial efficiency gains: financial platform Ramp reduced task runtime by 33% and costs by 58%, while framework developer LangChain slashed multi-turn agent costs by 74% with negligible impact on accuracy.
NVIDIA is collaborating on model routing integration with major ecosystem partners, including Siemens, Cadence Design Systems Inc., Cognition AI, and LangChain.
The release marks NVIDIA’s first open-source deployment since Huang joined industry executives, including Meta Platforms Inc. CEO Mark Zuckerberg, in advocating for open models before U.S. policymakers. National security concerns had spiked in Washington following the debut of China-based Moonshot AI’s Kimi K3 model, sparking discussions around potential sanctions and distillation restrictions.
Huang defended open-weight architectures in an open letter, arguing that accessible models accelerate cybersecurity, foster domestic innovation, and lower financial barriers. Beyond policy, the business rationale remains straightforward: reducing enterprise software costs encourages wider AI adoption, which in turn expands the hardware footprint required to run these workloads.
“Free AI should be great for hardware,” Huang said. “Free AI should be great for chips.”