NVIDIA Nemotron 3.5 Lightning
Fast, customizable open model for high-volume always-on agent workloads
About model
Nemotron 3.5 Lightning is NVIDIA's speed-focused open model for always-on agents, built to handle the high-volume, repetitive, domain-specific steps that make up most of an agent's work. A hybrid Mixture-of-Experts model with 30B total parameters and 3B active, distilled from NVIDIA's frontier Nemotron 3 Ultra and developed with the Nemotron Coalition, it pairs multi-token prediction and DFlash speculative decoding with a context window of up to 1M tokens. NVIDIA projects up to 4x higher throughput and up to 30% faster task completion on high-volume agentic workflows. Trained for popular agent harnesses and designed for post-training, so organizations can adapt it to their own tools, workflows, and policies for domain-specific accuracy. Available on Together AI.
Up to 4x
With up to 30% faster task completion on high-volume agent workflows
30B
Hybrid MoE keeping per-step inference cost at small-model levels
1M
Long-running, multi-turn agent sessions without losing state
- Fast Task Completion: NVIDIA-projected up to 4x higher throughput and up to 30% faster completion on high-volume, specialized agent steps
- Built for Agent Harnesses: Trained for popular agent harnesses, with reliability-focused behavior for multi-step workflows
- Designed for Customization: Open model trained on open datasets, built to be post-trained on an organization's own tools, workflows, and policies for domain accuracy
- Production-Ready Infrastructure: 99.9% SLA, available on Together AI dedicated infrastructure
Model card
Architecture Overview:
• Hybrid Mixture-of-Experts architecture with 30B total parameters and 3B active per token
• Multi-token prediction with DFlash speculative decoding for faster generation
• Context window of up to 1M tokens for long-running, multi-turn agent sessions
• Text input, text output
Training Methodology:
• Distilled from NVIDIA's frontier Nemotron 3 Ultra
• Developed with the Nemotron Coalition of AI labs and trained on open datasets
• Trained for agentic tasks and popular agent harnesses, targeting the high-volume specialized steps in multi-model agent systems
Performance Characteristics:
• NVIDIA-projected up to 4x higher throughput and up to 30% faster task completion on high-volume agentic workflows
• NVIDIA reports strong preliminary results on agent productivity and reliability benchmarks, including PinchBench and AA-Omniscience Non-Hallucination
Prompting
Together AI API Access:
• Access Nemotron 3.5 Lightning via Together AI APIs using the endpoint nvidia/nemotron-3.5-lightning
• Authenticate using your Together AI API key in request headers
• Text in, text out, with a context window of up to 1M tokens for long agent sessions
• Suited to serving the high-volume, specialized steps in multi-model agent systems
• Available on Together AI dedicated infrastructure
Applications & use cases
Always-On Agent Systems:
• Serve the high-volume, repetitive steps in agent pipelines while larger models handle frontier reasoning
• Run long-lived personal and productivity agents that manage email, calendars, projects, and bookings
• Keep multi-turn agent state across sessions with the 1M-token context window
Financial Services Workflows:
• Extract data from documents and prepare structured summaries at scale
• Check policy rules and monitor risk signals as dedicated agent steps
• Serve high request volumes with the model's 3B-active-parameter cost profile
Security Operations:
• Enrich alerts, classify incidents, and correlate indicators as always-on agent tasks
• Query logs and validate controls through structured agent workflows
• Prepare structured findings for analysts from continuous monitoring streams
Telecom & Retail Operations:
• Triage network alarms and answer billing questions with specialized agents
• Enrich product catalogs and resolve inventory and fulfillment exceptions
• Handle order, return, and loyalty questions at high volume
