NVIDIA Nemotron 3.5 Lightning is an open mixture-of-experts model from NVIDIA built for always-on AI agent workloads. It targets platform teams and enterprises that run high-volume agent loops and want a small, customizable model to handle routine tool calls instead of spending frontier-model tokens.
Core Features
- 30-billion-parameter MoE with only 3 billion active parameters per token, delivering up to 4x higher output speed than similar open models.
- Built for agent harnesses: it makes accurate tool calls with lower latency and shines on routine work like git pulls, output formatting, and CI checks.
- Context window up to 1 million tokens, shipped in BF16 and NVFP4 checkpoints with speculative decoding (MTP, DFlash, DSpark).
- Runs on Nvidia DGX Spark, RTX PRO workstations, Jetson edge devices, or data-center GPUs, so it deploys from a desk to a rack.
- Day-0 support across SGLang, OpenRouter, and Fireworks AI, plus NeMo Switchyard for routing requests to the right model.
Best For
- Platform teams standardizing reviews and delegating contained fixes across GitHub or GitLab.
- Enterprises customizing the model with NeMo on domain data for code review, security, or billing agents.
- Builders who need private, on-prem inference without sending every call to a frontier API.
Pricing
Nemotron 3.5 Lightning is open weight under NVIDIA's OpenMDW-1.1 license, which permits commercial use, fine-tuning, and self-hosting at no cost. There is no per-token fee to run it locally or on your own GPUs; NVIDIA also serves it through NIM and partner endpoints where standard inference pricing applies. The companion Nemotron-RL-Agentic-Terminal-Pivot dataset is released for post-training.
Our Take
Best for high-volume agent execution where you want frontier-level accuracy on cheap, local hardware. The trade-off is that it is an execution specialist, not a planner: a frontier model like Nemotron 3 Ultra or GPT-5.6 still needs to orchestrate the workflow. See more AI Engine/Model tools.
