NVIDIA has introduced new open-source technologies designed to power always-on AI agents composed of specialized models. The release includes two integrated components: Nemotron 3.5 Lightning, a lightweight, customizable open model optimized for high-volume agentic tasks, and NeMo Switchyard, an open-source routing library that dynamically directs each step of an agent workflow to the most capable and efficient model available.
These innovations address a significant structural challenge in multi-step AI agents: long-running agents devote most of their time to tool calls, result validation, and subagent delegation. Sending every one of these steps to a frontier reasoning model incurs unnecessary cost and latency. Nemotron 3.5 Lightning is a 30B mixture-of-experts (MoE) model with only 3B active parameters, built on a hybrid Mamba-2 + MoE + Attention architecture with a 1M-token context window. NVIDIA reports up to 4x faster output speed compared to similar-sized models, and a 30% faster completion rate on 10,000 PinchBench tasks than Qwen3.6 35B at comparable accuracy.
Industry players such as CrowdStrike, Harvey, CodeRabbit, Fastino Labs, and Lila Sciences are already customizing the model for cybersecurity, legal, coding, finance, and healthcare use cases.
Deployment Readiness
Yes. Nemotron 3.5 Lightning is generally available under the permissive OpenMDW-1.1 license, with open weights, training data, and recipes. NVIDIA confirms the model is ready for commercial deployment.
Key Features at a Glance
- Model architecture: Hybrid Mamba-2 + MoE + Attention, 30B total parameters, 3B active
- Context window: Up to 1M tokens
- Performance: 4x faster output than peers; 30% faster task completion on PinchBench
- License: OpenMDW-1.1, commercial-friendly
- Adoption: Verified by major enterprises across multiple verticals
What Is NeMo Switchyard?
NeMo Switchyard is an open routing engine that intelligently assigns each step of an agent pipeline—whether it involves generation, tool use, or verification—to the best-suited model. By combining a high-speed local model with selective use of larger remote models, Switchyard reduces cost and latency while maintaining reliability and quality.
Outlook for 2026
With the rapid expansion of agentic AI workloads in production, NVIDIA’s combination of an efficient MoE backbone and a smart routing layer positions Nemotron 3.5 Lightning as a practical building block for scalable AI systems. The emphasis on open weights and commercial usability is likely to accelerate adoption across industries seeking cost-effective, high-performance alternatives to monolithic models.
For developers and enterprises aiming to build responsive, budget-conscious AI agents, this release offers a compelling stack that balances performance, flexibility, and ease of integration.
via MarkTechPost