NVIDIA Nemotron 3.5 Lightning: Open MoE Built for Agents

NVIDIA's Open MoE for the Agent Era

On 11 August 2026, NVIDIA released Nemotron 3.5 Lightning — a 30-billion-parameter Mixture-of-Experts model with only 3 billion active parameters per forward pass and a 1-million-token context window. The model ships under the permissive OpenMDW-1.1 licence with full weights, training data, and recipes on Hugging Face and ModelScope. NVIDIA positioned the release explicitly for the execution layer of always-on agentic workloads: tasks that must run at high volume, low latency, and on hardware that organisations already own. Alongside the model, NVIDIA released NeMo Switchyard — a model router that orchestrates which model handles each step in a multi-agent pipeline.

Architecture: Hybrid Mamba-Transformer at 30B Scale

Nemotron 3.5 Lightning carries 31.6 billion total parameters but activates only 3.6 billion on any given inference call. This is the efficiency argument of Mixture-of-Experts: the model holds a large knowledge pool but the cost of each forward pass is proportional to active parameters, not total. The underlying architecture is a hybrid Mamba-Transformer design — the same lineage as NVIDIA's Nemotron 3 Nano — with substantial gains in intelligence and agentic task performance. On the Artificial Analysis Intelligence Index, Nemotron 3.5 Lightning scores 24, up from Nemotron 3 Nano's score of 15.

The 1-million-token context window matters for software engineering agents. An agent running code generation, test generation, or a multi-file refactor needs to hold the entire project in context without chunking — a context cliff that breaks reasoning chains and forces expensive re-reads. At 1M tokens, Nemotron 3.5 Lightning can hold a large codebase in a single context window, matching the working condition of real agentic software tasks rather than single-file code completion.

Benchmark Performance: Validated Against Real Tasks

NVIDIA published benchmark scores across knowledge and agentic dimensions. On SWE-bench Verified, the standard evaluation for software engineering agents running on real open-source GitHub repositories, Nemotron 3.5 Lightning scored 51.56. On GPQA Diamond — graduate-level reasoning across physics, chemistry, and biology — it scored 75.44. MMLU Pro came in at 81.94, and PinchBench, targeting agentic task efficiency, scored 85.37.

NVIDIA reports up to 4 times the output speed of similar-sized models, and 10,000 tasks completed 30 per cent faster than Qwen3-6-35B at comparable accuracy. Independent benchmarkers at Artificial Analysis have confirmed the 24-point Intelligence Index score — a 9-point gain over the predecessor model.

NeMo Switchyard: Routing the Right Task to the Right Model

Released alongside Nemotron 3.5 Lightning, NeMo Switchyard is NVIDIA's open model router for multi-agent pipelines. In a production agentic system, different subtasks carry different intelligence and latency requirements: a simple file-read or API call does not justify routing through a 70-billion-parameter frontier model, while a complex code synthesis step may require it. NeMo Switchyard provides the orchestration layer that sends each subtask to the appropriate model based on declared capabilities and current system load.

The practical result is cost reduction at the pipeline level. Organisations running always-on agents at large task volumes can mix Nemotron 3.5 Lightning — which fits on a single H100 or DGX Spark — with larger models for the subset of tasks that require them, without hard-coding model selection for each task type in application code.

Availability and Licensing

Nemotron 3.5 Lightning is available on Hugging Face and ModelScope as full-weight downloads, on OpenRouter for API access, and on build.nvidia.com as an NVIDIA NIM microservice. NVIDIA Cloud Partners, post-training platforms, and inference providers have broad coverage at launch. The OpenMDW-1.1 licence permits commercial use, fine-tuning, and redistribution — full weights, training data, and recipes included — placing it alongside Meta's Llama and MIT-licensed models in the genuinely open tier.

What This Means for Teams Building with AI in India

For engineering teams in India deploying AI agents, Nemotron 3.5 Lightning's single-H100 requirement and OpenMDW-1.1 licence matter practically. A model that fits on hardware available across most cloud environments — with no per-token API fees and no data sent to a third-party endpoint — lowers the deployment threshold for agentic systems handling sensitive data, a recurring concern for fintech and healthcare builders.

NVIDIA's simultaneous release of NeMo Switchyard signals the direction of the broader agent infrastructure stack: not monolithic frontier models, but routed pipelines where lightweight, fast models handle high-volume routine work and expensive models are reserved for tasks that genuinely require them. Indian software teams building at scale — in fintech, e-commerce, or enterprise SaaS — will increasingly evaluate their AI architectures in exactly these terms.

The Bottom Line

On 11 August 2026, NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter open Mixture-of-Experts model with 3 billion active parameters, a 1-million-token context window, and an OpenMDW-1.1 licence covering full weights, training data, and recipes. It scored 51.56 on SWE-bench Verified and runs on a single H100 or DGX Spark, delivering frontier-range agentic performance at a fraction of closed-model per-token costs. Alongside it, NVIDIA released NeMo Switchyard, a model router for directing subtasks within multi-agent pipelines to the most appropriate model at each step. Together they represent NVIDIA's architecture for production-scale always-on AI agents: fast, cheap specialists routing to powerful generalists only when required.