NVIDIA Nemotron 3.5 Lightning Release

NVIDIA Nemotron 3.5 Lightning is now available on Ollama, enabling the deployment of a 30 billion parameter open model designed specifically for local agentic workflows. This model is optimized for multi-step tasks—such as tool calling and context gathering—allowing developers to run high-performance agents on local hardware to ensure data privacy and reduce latency.

Model Architecture and Technical Specifications

Nemotron 3.5 Lightning utilizes a hybrid Mixture-of-Experts (MoE) architecture, featuring 30 billion total parameters but only 3 billion active parameters per token. This design allows the model to maintain the capabilities of a larger model while remaining efficient enough to run on local systems, including NVIDIA RTX PCs, RTX PRO workstations, DGX Spark, and DGX Station, as well as in cloud environments.

Key technical specifications include:

• Context Window: A context length of up to 1 million tokens, providing sufficient space for extensive tool histories in multi-turn workflows.
• Inference Optimization: The model employs speculative decoding via multi-token prediction (MTP), DFlash, or DSpark, which results in up to 4x higher throughput compared to similar open models.
• Customization: As an open model trained on open datasets, Nemotron 3.5 Lightning supports post-training for specialized tasks.

Agentic Capabilities and Use Cases

Nemotron 3.5 Lightning is specifically trained for agent harnesses, focusing on coding, tool calling, instruction following, and multi-turn work. It is optimized for workloads where agents must remain active to perform sequential operations like reading files, calling tools, and retrying failed steps.

Primary use cases include:

• Local Personal Assistants: Managing emails, calendars, and bookings using local context without transmitting data externally.
• Coding Sub-agents: Executing tests, searching codebases, and applying refactors within existing developer harnesses.
• ** uma Security Operations:** Enriching alerts, classifying incidents, querying logs, and preparing structured findings for analysts.
• Hybrid Cloud Tiers: Acting as a local tier that handles high-volume steps locally, while routing complex tasks to a larger hosted model via the same CLI and API.
• Specialized Local Models: Serving as a base for post-training on narrow, domain-specific jobs for local execution.

Performance Benchmarks

Nemotron 3.5 Lightning demonstrates significant efficiency gains over leading open models of a similar size. It offers 4x higher throughput and 30% faster task completion times. According to NVIDIA, higher throughput is critical for long-running agents because it increases the number of steps completed per minute, thereby reducing the total time required for complex tasks. The model also maintains leading accuracy across reasoning, coding, and agentic tasks.

Deployment via Ollama

Nemotron 3.5 Lightning can be deployed locally using Ollama. For users on Apple silicon, a specialized version (nemotron-3.5-lightning:30b-mlx) is available for optimized performance.

Users can run the model through general chat or integrate it with various tools such as Claude Code, OpenClaw, Hermes Agent, and OpenCode using the ollama launch command.