NVIDIA has introduced Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts (MoE) open model designed for high-volume execution tasks in long-running AI agents. The company has also released NeMo Switchyard, an open-source model routing library that directs requests to different AI models based on requirements such as quality, latency and cost.
Nemotron 3.5 Lightning is designed for specialized tasks within multi-agent systems, while larger reasoning models can handle planning and orchestration. NVIDIA says it delivers up to 4x faster output than similar-sized models and enables 30% faster agentic task completion.
Always-on agents and systems of models
Long-running AI agents increasingly use multiple models for different tasks. Larger reasoning models such as Nemotron 3 Ultra can handle planning and orchestration, while smaller models such as Nemotron 3.5 Lightning can perform tasks including code review, tool use, security alert monitoring and answering billing questions.
NVIDIA Nemotron 3.5 Lightning
Nemotron 3.5 Lightning is the smallest model in the Nemotron 3 family, with 30 billion total parameters and 3 billion active parameters. Its MoE architecture activates only a subset of experts for each token.
The model is designed for high-volume agent execution, including tool calls, result validation and subagent delegation. NVIDIA said it was developed with contributions from the Nemotron Coalition, which provided evaluation methodologies, inference software and datasets.
Key specifications:
- Architecture: 30B MoE
- Active parameters: 3B
- Formats: BF16, NVFP4
- Primary use: High-volume agent execution
- Deployment: Local, on-premises, edge, workstation, data center and cloud
Performance and optimization
NVIDIA says Nemotron 3.5 Lightning provides up to 4x faster output than other models in its class and completes agentic tasks 30% faster. On PinchBench, it reaches 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy.
The model uses multi-token prediction (MTP) and speculative decoding. It includes two draft models:
- DSpark: Recommended for DGX Spark inference and low-concurrency data center workloads.
- DFlash: An additional draft model for workload evaluation.
MTP is intended for medium- to high-concurrency workloads, with the optimal draft length decreasing as concurrency increases.
The NVFP4 checkpoint uses specialized NVFP4 kernels also used by Nemotron 3 Ultra and supports NVIDIA Blackwell, Hopper and Ampere GPUs. The model also uses harness-optimized training for popular agent harnesses.
Customization and deployment
Nemotron 3.5 Lightning can be customized with an organization’s own data, tools and workflows. NVIDIA has released its weights, training data and recipes as permissively as licensing permits under OpenMDW-1.1.
Developers can use NeMo Automodel and NeMo Megatron Bridge for LoRA or full SFT, while NeMo RL and NeMo Gym support reinforcement learning, evaluations and rollouts.
The release also includes Nemotron-RL-Agentic-Terminal-Pivot, an open reinforcement learning dataset used for coding-agent capabilities.
NVIDIA said it publishes as much of the training data and techniques as licensing permits for traceability, auditing and training other models.
Industry use cases
NVIDIA said companies are customizing the model for:
- CrowdStrike: Cybersecurity
- Harvey with Trajectory: Legal services
- CodeRabbit with Baseten: Code review
- Lila Sciences: Physical and life sciences
- Fastino Labs: Software development, finance and healthcare
Nemotron 3.5 Lightning can run on NVIDIA RTX PCs, DGX Spark, DGX Station, Jetson, GeForce RTX 5090 and RTX PRO workstations. It is also supported by LM Studio, llama.cpp, Ollama and Unsloth.
NVIDIA said it worked with teams including EXO Labs to evaluate the model on DGX Spark.
The model supports agent harnesses including OpenClaw and Hermes Agent, which are supported by NVIDIA’s NemoClaw open-source security and management stack.
Accuracy and efficiency
Nemotron 3.5 Lightning reaches the accuracy-speed Pareto frontier on the Artificial Analysis Intelligence Index, according to NVIDIA. The index combines nine evaluations covering agentic tasks, coding, scientific reasoning and general intelligence.
On PinchBench, the model reaches 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy. NVIDIA says its inference throughput and token efficiency place it on the efficiency frontier for high-volume workloads.
NeMo Switchyard
NeMo Switchyard is an open-source model routing library for AI agents. It can route prompts to different models at each step of an agent workflow based on requirements such as quality, latency and cost.
Nemotron 3.5 Lightning can be used as a routing target alongside other open and closed models, allowing larger models to handle planning while smaller models handle execution.
NVIDIA’s internal benchmarks show that Switchyard maintains frontier-level accuracy while reducing task completion cost to nearly one-third of using Opus 4.8 alone.
Companies including Boomi, Cadence, Classmethod, Cognition, Kong, LangChain, LiteLLM, Nous Research, Ramp and Siemens are evaluating or integrating Switchyard across AI tools and platforms.
Partner ecosystem
Nemotron 3.5 Lightning is supported across post-training, inference, agent frameworks and cloud platforms.
- Post-training: AgileRL, Applied Compute, Deep Cogito, Fastino Labs, Locai Labs, Prime Intellect, Reasonable, Thinking Machines Lab, Thoughtworks, Trajectory, Uniphore
- Inference: Ollama, Exo, Canonical, LM Studio, Unsloth
- Agent frameworks: Aible, Cline, Factory AI, Hermes Agent, Kilo Code, LangChain, OpenClaw, OpenCode, OpenHands, Pi
- Cloud platforms: Amazon SageMaker JumpStart, Google Cloud Gemini Enterprise Agent Platform, MSFT Foundry, OCI Enterprise AI
- GSI: Accenture, Tata Consultancy Services, Tech Mahindra, Wipro
- Hosted inference: Baseten, CoreWeave, Crusoe, DeepInfra, Fireworks AI, FriendliAI, GMI Cloud, Modal, Nebius, Together AI
Availability
Nemotron 3.5 Lightning is available through Hugging Face, ModelScope, OpenRouter and build.nvidia.com, including as an NVIDIA NIM microservice. It is also available through NVIDIA Cloud Partners and other inference, post-training and cloud platforms.
NeMo Switchyard is available on GitHub and is coming to partner platforms. NVIDIA also provides deployment guides for vLLM, SGLang and TensorRT-LLM, along with documentation for Nemotron 3.5 Lightning and Switchyard.