Overview
Run NVIDIA Nemotron 3.5 Lightning 30B A3B on Crusoe. NVIDIA’s efficient hybrid MoE for long-running agents, ready to serve, deploy, or fine-tune.
Explore additional models
Frequently
asked questions
Nemotron 3.5 Lightning 30B A3B is NVIDIA’s newest efficiency model, released in August 2026: a 32.9B hybrid Mamba-Transformer mixture-of-experts model with about 3B active parameters, built for long-running agentic workloads. On Crusoe it supports a 262,144-token context and is available for Serverless Inference, Self-Serve Deployments and Serverless Fine-Tuning, with no infrastructure setup required.
Nemotron 3.5 Lightning 30B A3B is the next generation of Nemotron 3 Nano 30B A3B in the same size class. Both have roughly 30B total parameters, about 3B active, and a hybrid Mamba-Transformer design. Lightning improves reasoning and instruction following for multi-step agent work. On Crusoe, Lightning is also available for Self-Serve Deployments and Serverless Fine-Tuning, while Nano is available for Serverless Inference only.
Nemotron 3.5 Lightning 30B A3B is billed based on the option you choose. Serverless Inference is billed per token, with separate rates for uncached, cached, and output tokens, and no minimum commitment. Self-Serve Deployments are billed per GPU hour, and Serverless Fine-Tuning is billed per 1M training tokens. Get started quickly with Crusoe’s OpenAI-compatible API in Crusoe Intelligence Foundry. Current rates are shown in the pricing section above.
NVIDIA reports Nemotron 3.5 Lightning 30B A3B running about 30% faster than Qwen3.6 35B A3B on PinchBench with comparable accuracy. NVIDIA positions it for specialized, long-running agent tasks where sustained accuracy across many steps matters more than single-turn scores.
Yes. Run Nemotron 3.5 Lightning on Serverless Inference with no infrastructure to manage. Get a dedicated endpoint optimized for throughput or responsiveness with Self-Serve Deployments. Or customize the model with your own data in a few steps using Serverless Fine-Tuning. Upload a JSONL or Parquet dataset, run a LoRA job with pre-configured defaults, then deploy the result in one click or download the weights. Because only about 3B parameters are active per token, the fine-tuned result serves efficiently on a single GPU.
“Lightning” is NVIDIA’s name for its speed-focused Nemotron model, not a technical acronym. It signals what the model is built for: fast, high-volume execution of the routine steps inside long-running agents, like tool calls, extraction, and task handoffs. NVIDIA reports Lightning delivering up to four times the output speed of comparable open models, which is the speed the name refers to.


%204.png)
