Nvidia’s Nemotron 3.5 Lightning: Small, Fast, and Open by Design

Nvidia has released Nemotron 3.5 Lightning, a new open-weight AI model that deliberately doesn’t try to compete with the biggest and smartest systems out there. With only 3.6 billion of its 31.6 billion total parameters active at any given time, it barely matches the intelligence score of OpenAI’s considerably older gpt-oss-120b on Artificial Analysis’s ranking. In exchange, Nvidia says it’s the fastest model in the entire comparison field, and that shift in priorities says a lot about where the market for open AI models is currently heading.

Key takeaways

• Nemotron 3.5 Lightning has 31.6 billion parameters total, but only 3.6 billion are active for any given request.
• On Artificial Analysis’s Intelligence Index it scores 24, tying OpenAI’s gpt-oss-120b and trailing Nvidia’s own, four-times-larger Nemotron 3 Super (26).
• At roughly 669 tokens per second, it’s the fastest model tested, nearly twice as fast as Google’s Gemini 3.5 Flash-Lite.
• On agentic benchmarks (GDPval-AA v2) it actually beats larger models, including gpt-oss-120b and Nvidia’s own Nemotron 3 Super.
• The model is available immediately under the open OpenMDW-1.1 license via Hugging Face and several cloud providers.

Built small, trained on purpose

Technically, Nemotron 3.5 Lightning relies on a hybrid architecture combining Mamba-2 and Mixture-of-Experts layers with a handful of classic attention layers. That design lets it activate only a fraction of its 31.6 billion parameters for any single request, which drastically cuts compute and memory needs. Nvidia says the context window scales up to one million tokens, and speculative decoding (DFlash) plus multi-token prediction further shorten response times. Compared with its own predecessor, the Nemotron 3 Nano that Lightning replaces, the jump on the Intelligence Index from 15 to 24 points is substantial, even though rivals like Qwen3.6 35B A3B (32 points) or Meta’s open Muse Glimmer, which kabel-salat.info covered earlier, are still clearly ahead.

Speed is the actual selling point

The real story here isn’t raw intelligence, it’s speed. At 669 tokens per second, Nvidia claims up to four times higher throughput and roughly 30 percent faster task completion in heavily automated, agent-based workflows. For exactly this kind of scenario, where an agent works through many small, specialized steps in sequence rather than answering one complex question, Nvidia is positioning the model as an “efficient workhorse.” According to cloud partners, the 30-billion-parameter version can even run on a single H100 GPU, which meaningfully cuts operating costs for companies that need many parallel agent instances rather than one giant model. Lightning was available on launch day through providers including DeepInfra, Fireworks, Together AI, OpenRouter, and CoreWeave. For development teams, that translates into a practical option: instead of keeping one large model on standby for every request, the same hardware budget can run several specialized Lightning instances in parallel, say one each for research, code execution, and summarization within a single agent workflow.

Why Nvidia releases open models at all

The real punchline is the business model behind it. Nvidia doesn’t make its money from AI models, it makes money from the chips those models run on, and every additional open model that companies can download for free and run themselves increases demand for exactly that hardware. That makes Nvidia one of the few major US companies still releasing open weights at meaningful scale, at a time when cheap Chinese model families like Qwen, DeepSeek, GLM, Kimi, and MiniMax keep closing the gap to closed frontier systems from Anthropic and OpenAI. Nvidia apparently isn’t just answering that pressure with compact, fast models like Lightning, insiders say the company is already working on a far larger Nemotron 4 with at least a trillion parameters, expected to be ready by late autumn and explicitly meant to compete with the world’s best open models.

Key takeaways for the road ahead

Nemotron 3.5 Lightning isn’t a model for anyone chasing the smartest system available, it’s for anyone who wants to automate a lot of simple tasks quickly and cheaply. Nvidia deliberately not aiming for maximum intelligence is itself a statement: competition among open models is no longer just about benchmark scores, it’s about speed, operating cost, and how much GPU capacity a model actually burns in day-to-day use. Companies planning to run their own AI agents would do well to take a second look at models like this one before reflexively reaching for the biggest name available. Whether this strategy actually sells more GPUs than a single giant model would is hard to prove from the outside, but the underlying bet is straightforward: the more open models run in everyday use, the more data centers get built to host them, and the chips inside those data centers will most likely keep coming from Nvidia.