project
Nemotron 3.5 Lightning - NVIDIA's open-source MoE model
Nemotron 3.5 Lightning is an open-source MoE model with 30 parameters from NVIDIA, optimized for multi-agent systems. Compared to similar models, it offers 4x faster output speed and 30% faster agent task completion...
What is Nemotron 3.5 Lightning?
Nemotron 3.5 Lightning is an open-source 30-parameter MoE model from NVIDIA, optimized for multi-agent systems. Compared to similar models, it offers 4x faster output speed and 30% faster agent task completion, achieving an excellent balance between accuracy and speed. The model supports deployment across all scenarios, from RTX PCs and Jetson to data centers, and is open-sourced on platforms such as Hugging Face and NVIDIA Build, facilitating the development of more efficient agent applications.
Main functions of Nemotron 3.5 Lightning
- Multi-agent task executionThis model is specifically designed for specific tasks in large-scale multi-agent systems, helping developers build smarter and more efficient agent applications.
- High-speed inference outputCompared to similar models, Nemotron 3.5 Lightning achieves up to 4x faster output speed, accelerating the overall completion speed of agent tasks by 30%.
- Flexible deployment across all scenariosThe model can run directly on local AI systems (including NVIDIA RTX PCs, DGX Spark, Jetson, etc.) and can also be seamlessly extended to edge devices, workstations, data centers and cloud environments.
The technical principles of Nemotron 3.5 Lightning
- MoE Hybrid Expert ArchitectureNemotron 3.5 Lightning employs a 30B parameter MoE (Hybrid Expert) architecture with a total of 30 billion parameters. Each inference only activates a subnetwork with approximately 3 billion parameters. Through a sparse activation mechanism, it significantly reduces computational overhead while maintaining the capabilities of large models, achieving efficient inference.
- NVFP4 Quantization CompressionThe model supports the NVFP4 quantization format, which is NVIDIA's self-developed 4-bit floating-point precision solution. It can significantly compress the model size and memory usage, enabling it to run smoothly on consumer and edge devices such as RTX PCs and Jetson.
- Intelligent agent scenario-specific optimizationThe model is specifically trained and optimized for multi-agent collaboration scenarios, focusing on improving the stability and accuracy of key aspects such as long-chain inference, tool invocation, and task decomposition, so as to better support complex agent workflows.
- Speed-accuracy joint optimizationIn terms of architecture design and training strategy, Nemotron 3.5 Lightning achieves joint optimization of accuracy and inference speed. It demonstrates superior overall performance compared to similar models in the PinchBench benchmark test, ensuring the quality of task completion and improving the overall execution speed of the agent by 30%.
Follow us on WeChat and reply with "open source",join inAI open source project discussion group
How to use Nemotron 3.5 Lightning
- Hugging Face DownloadAccess the Hugging Face official repository to download the model weights, which can then be loaded and run locally using Nemotron 3.5 Lightning.
- NVIDIA Build platformVisit the NVIDIA Build website for model introductions, technical documentation, and deployment guidelines.
- Local device deploymentDeploy the model directly on local devices such as NVIDIA RTX PC, DGX Spark, DGX Station, or Jetson.
- Enterprise-level expansionSeamlessly expands to RTX PRO workstations, data centers, or cloud environments to meet enterprise-level application needs.
The core advantages of Nemotron 3.5 Lightning
- MoE Sparse Activation ArchitectureThe model adopts a MoE architecture with 30B total parameters and approximately 3B activation parameters, which significantly reduces computational overhead while maintaining the capabilities of a large model.
- Extremely fast reasoning performanceCompared to similar models, it increases output speed by 4 times and can accelerate the overall completion speed of agent tasks by 30%.
- NVFP4 Quantization Support: By leveraging NVIDIA's self-developed 4-bit floating-point quantization format, it significantly compresses video memory usage, enabling deployment across all scenarios from RTX PCs to data centers.
- Intelligent Agent Specific OptimizationIt is specifically trained for long-running multi-agent tasks, achieving an excellent balance between accuracy and inference speed.
Nemotron 3.5 Lightning project address
- Project official website：https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightning-delivers-fast-accurate-specialized-task-execution-for-long-running-agents/
- HuggingFace model library：https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
Nemotron 3.5 Lightning vs. Competitors
| Comparison Dimensions | Nemotron 3.5 Lightning | Qwen3.6-35B-A3B |
|---|---|---|
| Publisher | NVIDIA | Ali Tongyi 1000 Questions |
| Total number of parameters | 30B | 35B |
| Activation parameter quantity | Approximately 3B (30B-A3B) | 3B |
| Architecture | MoE (Hybrid Expert) | MoE (Hybrid Expert) |
| Context length | Not explicitly mentioned | 200K |
| Open source license | Fully open source | Apache 2.0 (commercially usable) |
| Reasoning speed | Output speed increased by 4 times, task completion speed increased by 30%. | Efficient reasoning, specific multiple not disclosed. |
| Programming skills | No specific benchmark was specified. | SWE-bench Verified 73.4，Terminal-Bench 2.0 51.5 |
| Multimodal support | Unclear | Native support, with visual capabilities on par with Claude Sonnet 4.5. |
| Reasoning patterns | Unclear | Supports flexible switching between thinking and non-thinking modes. |
Application scenarios of Nemotron 3.5 Lightning
- Multi-agent cooperative systemAs a dedicated task model, it drives complex agent workflows, efficiently completing tool calls, task decomposition, and long-chain inference.
- Local edge AI inferenceLeveraging NVFP4 quantization and sparse activation mechanisms, it enables low-latency, low-power local inference on RTX PCs and Jetson devices.
- Enterprise-level intelligent agent platformSeamlessly expands to data centers and the cloud, supporting high-frequency, long-term enterprise-level intelligent agent services and automated processes.
- Code Development and DevOps AutomationIt enables rapid generation, review, and debugging of code, accelerating the execution of agent tasks in software development and continuous integration.
- Real-time data analysis and decision supportIn business scenarios requiring rapid response, it provides high-speed and accurate reasoning results to assist in real-time decision-making.