Nvidia released Nemotron 3.5 Lightning on August 11. It's a 30-billion-parameter open-weight model. Only 3 billion of those parameters activate at each inference step — the model routes each token through specialist sub-networks rather than running all 30 billion every time. That's the Mixture of Experts part. That's also why it fits on a single H100-80GB at 256K context lengths, where a dense 30B model wouldn't.

Artificial Analysis benchmarked it at 24 on the Intelligence Index, up from 15 for its predecessor Nemotron 3 Nano 30B, and found it performs comparably to gpt-oss-120b — OpenAI's 120-billion-parameter open model — at roughly a quarter of the total parameters. For agentic workloads: 4× the output token speed of the prior Nemotron generation, and 10,000 tasks on Nvidia's PinchBench agentic eval completed 30% faster than Qwen3.6 35B. Jensen Huang has been saying publicly that open-weight AI should be competitive with closed models. This is the most substantive product evidence Nvidia has shipped for that thesis.

Source spread

• Nvidia Technical Blog — Nemotron 3.5 Lightning: fast, accurate specialized task execution [hype] — Primary source; all benchmark claims and hardware specs originate here.
• Artificial Analysis — Nemotron 3.5 Lightning launch [builder] — Intelligence Index score and gpt-oss-120b comparison, independent of Nvidia.
• The Decoder — Nemotron 3.5 Lightning prioritizes speed over maximum intelligence [skeptic] — Frames this correctly as an efficiency-first model, not a frontier-intelligence release.
• Hugging Face — nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 [builder] — Weights, model card, license text, hardware requirements.

Pros & cons

What's real:

• The MoE efficiency gain is genuine. Paying inference compute on 3B active parameters instead of 30B is a real cost reduction, not a marketing reframe. The routing mechanism is doing actual work.
• The license is meaningfully open. OpenMDW-1.1 includes weights, training data, and training recipes. That's a different category from "we released weights" — you can study the distillation pipeline and adapt it. Most "open" releases don't go this far.
• Single H100-80GB deployment at 256K context is useful for builders who have one or two GPUs and want local, offline inference without an API dependency. DGX Spark (the GB10 consumer chip) also works.
• NeMo Switchyard ships alongside it. If you're building multi-model agent pipelines and routing calls by task type, a router trained to know when Lightning is sufficient versus when you need something larger is the practical complement. Worth evaluating as a stack, not individually.

What deserves a side-eye:

• The Intelligence Index is Artificial Analysis's composite benchmark, not SWE-bench Verified or TAU-bench. Composite scores can hide task-specific gaps. If your workload is heavy on tool use at long context — TAU-bench territory — verify there before drawing production conclusions from the composite.
• "Distilled from Nemotron 3 Ultra" means the intelligence ceiling is bounded by the teacher. This is not a frontier reasoning model. Don't evaluate it as one. Use it for the fast, high-volume decisions in an agent stack — not the final step where being wrong carries real consequences.
• The 1M-token context window needs 8× H100 or GB200. The single-H100 story is for 256K context. That caveat is buried in the hardware requirements and doesn't appear prominently in the headline coverage.
• OpenMDW-1.1 is Nvidia's own license, not Apache 2.0 or MIT. It's permissive — but worth reading the actual document before relying on it for anything production-critical.
• PinchBench is Nvidia's internal agentic benchmark. The 30% speed claim against Qwen3.6 35B is first-party. Verify against your own workload before using it for capacity planning.

Samwise's take

What builders need to know

Further reading

• Nvidia Technical Blog — Nemotron 3.5 Lightning — primary source, all benchmark claims
• Hugging Face — nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 — weights, model card, license
• Artificial Analysis — Nemotron 3.5 Lightning launch — Intelligence Index comparison and gpt-oss-120b benchmark
• MarkTechPost — Nvidia AI Releases Nemotron 3.5 Lightning and NeMo Switchyard — NeMo Switchyard router details
• Nvidia NGC — Nemotron 3.5 Lightning model card — hardware requirements and deployment options