Nvidia's New Free Model Is Not the Smartest. It Is Just Very, Very Fast
Nemotron 3.5 Lightning matches OpenAI's gpt-oss-120b on intelligence scores with a quarter of the parameters, and produces around 670 tokens per second, the fastest in its comparison group.
Nvidia released Nemotron 3.5 Lightning on Tuesday, an open-weights model you can download and run yourself. It is not trying to be the smartest model in the room. It is trying to be the quickest, and by that measure it currently wins its class.
The numbers, via the independent benchmarking site Artificial Analysis: Lightning scores 24 on their Intelligence Index, a nine point jump over its predecessor and exactly level with OpenAI’s gpt-oss-120b, a model roughly four times its size. Where it pulls away is speed. Lightning produces close to 670 tokens per second, which is the highest measured throughput in the whole comparison and almost twice as fast as Google’s Gemini 3.5 Flash-Lite at 386. A typical benchmark task takes it about half a minute. Qwen3.6 35B A3B needs around 3.5 minutes for the same work, and Gemma 4 31B roughly 5.8.
The trick is in the architecture. Lightning has 31.6 billion parameters in total, but only 3.6 billion of them are switched on for any given step. Parameters are the adjustable numbers a model learns during training, and this “mixture of experts” design means the model carries a large amount of knowledge while only paying the computing cost of a small model each time it answers. It handles text only, reads up to one million tokens of context at once, and ships under the permissive OpenMDW-1.1 licence.
What is behind it
Nvidia has been arguing this case for over a year. In a widely discussed paper, its researchers claimed that models under 10 billion parameters could handle most agent workloads about as well as models ten to thirty times larger, at a fraction of the cost. Lightning is the clearest product built on that argument so far. And the agentic benchmarks back it up: on GDPval-AA v2 it hits an Elo rating of 824, beating both gpt-oss-120b at 800 and Nvidia’s own, larger Nemotron 3 Super at 698.
A fair caveat, though. Smarter small models exist. Qwen3.6 35B A3B scores 32 and Meta’s Muse Glimmer 35, both well ahead of Lightning’s 24. Proprietary models are further ahead still. Lightning is built for one job: doing many small steps quickly and cheaply, which is exactly what an AI agent working through a long task does all day.
What this means for you: If you have never run a model yourself, nothing changes today, and that is fine. If you are already experimenting, this is worth a look, because speed is the thing that makes agents feel usable rather than sluggish. The weights are on Hugging Face, and several providers including DeepInfra, Fireworks, Nebius and CoreWeave already serve it if you would rather not host it yourself. Just do not expect it to reason like a frontier model. That is not what it is for.
Sources
Researchers Found a Way to Read the Hidden Thoughts of Claude, ChatGPT and Gemini
A paper published on stolen-thoughts.com shows how encrypted reasoning blocks from frontier models could be replayed into weaker sibling models and decoded. All three providers have now fixed it.
