Nvidia has released Nemotron 3.5 Lightning, the first model in a new Nemotron 3.5 family, and it is playing a different game than most open-weight releases this year. Instead of chasing the highest intelligence score, Nvidia built a compact model that runs fast and cheap while still matching much larger rivals on standard benchmarks.
A small model that keeps up
Lightning carries 31.6 billion total parameters, but only 3.6 billion are active at any moment. It keeps the hybrid Mamba-Transformer design of its predecessor, Nemotron 3 Nano 30B A3B, and ships as an open-weights model under the permissive OpenMDW-1.1 license.
On the Artificial Analysis Intelligence Index, Lightning scores 24. That is a nine-point jump from the Nano it replaces, and it ties OpenAI's gpt-oss-120b, a model with roughly four times the active footprint. Nvidia's own larger Nemotron 3 Super scores 26. The clearest leads in the small-model class still belong to Qwen3.6 35B A3B at 32 and Meta's Muse Glimmer at 35.
Where it wins: throughput
The headline number is speed. In pre-release tests using the final NVFP4 weights, Lightning sustained about 669 tokens per second, the highest measured throughput in the comparison and nearly twice Google's Gemini 3.5 Flash-Lite at 386 tokens per second.
That throughput shows up in wall-clock time. A task drawn from the Intelligence Index finished in about 0.5 minutes on Lightning, against roughly 3.5 minutes for Qwen3.6 35B A3B and about 5.8 minutes for Gemma 4 31B. For agent pipelines that call a model many times, that gap is the difference between a workflow that feels instant and one that feels like a queue.
Agentic benchmarks favor it
The biggest gains land on agentic evaluations. On GDPval-AA v2, Lightning reaches an Elo rating of 824, up 334 points from Nano, and ahead of both gpt-oss-120b at 800 and the larger Nemotron 3 Super at 698. On Terminal-Bench v2.1 it climbs from 7 percent to 24.3 percent, close to gpt-oss-120b's 26.2 percent.
Nvidia says post-training work with partners including CodeRabbit and Harvey pushed performance in specific domains, positioning the model as a high-throughput workhorse for agent-based systems rather than a general frontier contender.
The tradeoff
Lightning is not trying to top the intelligence charts. Proprietary models still own the efficiency frontier on raw score: Gemini 3.5 Flash-Lite posts 37 on the Intelligence Index in a similar time per task, and GPT-5.6 Luna (max) reaches 52 in under two minutes.
What Nvidia is offering instead is a self-hostable, open-weights option that does the bulk of the work at a fraction of the compute bill. For teams shipping agents at scale, that is often the constraint that matters most.
Sources
Nvidia's open-weight Nemotron 3.5 Lightning prioritizes speed over maximum intelligence - The Decoder, August 11, 2026: https://the-decoder.com/nvidias-open-weight-nemotron-3-5-lightning-prioritizes-speed-over-maximum-intelligence/
Nemotron 3.5 Lightning benchmark data - Artificial Analysis: https://artificialanalysis.ai/models/nemotron-3-5-lightning
Nemotron 3 Nano swaps pure Transformers for a Mamba hybrid - The Decoder: https://the-decoder.com/nvidias-nemotron-3-swaps-pure-transformers-for-a-mamba-hybrid-to-run-ai-agents-efficiently/
