Nvidia Ships Nemotron 3.5 Lightning, a 30B Open Agent Model Distilled From Its Own Ultra Flagship
Nvidia released Nemotron 3.5 Lightning, a 30B mixture-of-experts model with only 3B active parameters built for high-volume agent workloads. It runs 30% faster than Qwen3.6 35B on Nvidia's own PinchBench agent suite.
Nvidia released Nemotron 3.5 Lightning on August 11, a 30-billion-parameter open model with just 3 billion active parameters per token, distilled from its larger Nemotron 3 Ultra. The target isn’t chatbots — it’s long-running agent workloads that fire thousands of tool calls per session.
On Nvidia’s own PinchBench agent suite, Lightning finished a 10,000-task run 30% faster than Alibaba’s Qwen3.6 35B, while staying competitive on knowledge, reasoning, instruction-following, coding, and long-context evals. Nvidia published its evaluation recipes in NeMo Gym so outside labs can reproduce the numbers rather than take the benchmark on faith — a direct answer to the industry’s chronic benchmark-gaming problem.
The mixture-of-experts architecture is the point. A 30B model with only 3B active parameters costs a fraction of a dense 30B model to serve, which is why Nvidia is positioning Lightning as the default worker model behind NeMo Switchyard, a router it launched alongside it. Switchyard triages incoming agent tasks and hands the easy, high-volume ones to Lightning while routing anything that needs deeper reasoning up to Nemotron 3 Ultra. That’s a deliberate bet: most enterprise agent traffic is repetitive tool orchestration, not frontier reasoning, so paying flagship-model prices for every step is wasteful.
Licensing matters here too. Nemotron 3.5 Lightning ships under Nvidia’s OpenMDW-1.1 license, which permits commercial use, and it’s already live on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as an NVIDIA NIM microservice. Launch-day OpenRouter pricing landed at $0.10 per million input tokens and $0.25 per million output tokens — cheap enough to run as the default tier for high-frequency agent pipelines rather than a specialty option.
The release lands a week after Nvidia helped stand up the 37-member Open Secure AI Alliance, and it reinforces the same strategy: Nvidia is positioning itself as the infrastructure and open-model layer underneath the agent economy, not just the chip vendor selling into it. With Meta’s Muse Glimmer and now Nemotron 3.5 Lightning both shipping as compact, single-GPU-friendly open agent models within days of each other, the market for lightweight local agent workers is getting genuinely competitive — good news for developers who don’t want to pay frontier-model rates to have an agent check a calendar or fill a form.
For teams building agent pipelines, Lightning is worth benchmarking against whatever is currently doing tool-call triage — the active-parameter math alone should cut inference cost meaningfully if the task mix skews toward routine execution rather than novel reasoning.