Thirty billion parameters, and only three billion of them wake up per token. That ratio is the whole pitch for what NVIDIA published on 11 August: Nemotron 3.5 Lightning, a mixture-of-experts model with open weights on Hugging Face under a plain OpenMDW 1.1 licence, aimed at agents that run all day instead of chatbots that answer once. The speed claim is the loud part, and unusually it survives an outside harness. Artificial Analysis clocked it at 350.9 output tokens per second against a median of 99.5 for models in its class. What doesn't survive is the context number, and one cost comparison NVIDIA really shouldn't have led with.
The short answer
NVIDIA published Nemotron 3.5 Lightning on 11 August 2026: a 30 billion parameter MoE with 3 billion active, on a hybrid Mamba-2 and attention stack, under the permissive OpenMDW 1.1 licence with NVFP4 and BF16 checkpoints. The agentic post-training dataset shipped with it. Independent measurement backs the speed claim. The advertised context and the served context don”t match, and the cost comparison NVIDIA leads with is its own benchmark.
What the sparsity actually buys you
Ten to one. That’s the ratio between what the model weighs and what it computes, and it’s why this thing is fast rather than merely cheap to store.
A mixture-of-experts model keeps all 30 billion parameters resident but routes each token through a small slice of them. So memory scales with the full model while compute scales with the 3 billion. NVIDIA went further and built the backbone as a hybrid, mixing Mamba-2 state-space layers with attention and the MoE blocks, which is where the throughput comes from on long inputs. Attention costs grow with sequence length. Mamba layers don’t, or rather they grow far more politely.
The result reads well on an outside harness. Artificial Analysis scores it 24 on their Intelligence Index against a median of 9 for comparable open-weights models, at 350.9 output tokens per second against a median of 99.5, with a 0.85 second time to first token where the median is 1.99.
Image: NVIDIA
I’ll admit I expected the independent numbers to deflate the launch post. They didn’t, and that’s rare enough to say plainly.
The part that shipped quietly
Weights under OpenMDW 1.1, fine. Plenty of that about lately.
The dataset is the interesting bit. NVIDIA released Nemotron-RL-Agentic-Terminal-Pivot, the agentic reinforcement learning set used to post-train the model for coding agent behaviour, which means the recipe for the specific capability being marketed is sitting there for anyone to run against their own base model. Weights without data are a product. Weights with the post-training data are closer to a method. That distinction matters more than the licence text, honestly, and it lands in the middle of a policy argument we wrote up when twenty-five signatories put their names to an open-weights letter.
Where the release stops being straight with you
Start with context. The model card says up to 1M tokens. OpenRouter serves 262K. Nobody reconciles those, and a factor of four is not a rounding difference. If you’re building around long documents, assume 262K until your own testing says otherwise.
Then the hardware framing. NVIDIA’s blog says RTX PCs, DGX Spark, DGX Station and Jetson. The model card, more carefully, names one H100 or one DGX Spark as the single-GPU deployment targets, and lists GeForce RTX 5090 among supported hardware without publishing a memory floor for it. Those aren’t contradictory, they’re just pitched at different audiences. A 5090 is a 32 GB card and a 30B model in NVFP4 should fit, but should isn’t a spec sheet. Compare that with the 24 GB floor Meta published for Muse Glimmer two days earlier, where at least the number was stated and you could plan against it.
And then Switchyard. The routing library itself is a genuinely useful piece of plumbing, an open-source layer that picks a model per request so you’re not paying frontier rates for tasks that don’t need them. The claim attached to it is where I’d push back: internal benchmarks showing task completion cost at nearly one third of Opus 4.8 alone. NVIDIA built the router, chose the workload, picked the comparison model, and ran the test. Any of us would produce a favourable number under those conditions. Nothing suggests it’s wrong. It’s just not evidence yet.
Who should pick this up
If you’re running high-volume agent loops where the same task shape repeats thousands of times, the economics here are hard to argue with. Half a dollar buys ten million output tokens, the latency is low enough for tool-calling chains that pause constantly, and SWE-bench Verified at 52.80 on the NVFP4 checkpoint says it can hold a coding task without being a frontier model about it.
If you want one strong model to think hard about a small number of problems, this isn’t it, and NVIDIA doesn’t claim it is. Lightning is explicitly the execution layer, the thing Switchyard routes the easy work to.
Two limits before you commit. Post-training data stops in May 2026 and pre-training in September 2025, so recency arrives through tools or not at all. And the language list is short: English plus code, Spanish, French, German, Italian, Japanese. If your users write in something else, test before you promise anything.
Sources
NVIDIA’s announcement is on the NVIDIA blog, with the technical write-up on the NVIDIA developer blog. Architecture, licence, hardware targets and the benchmark table come from the model card on Hugging Face. Independent speed and intelligence measurements are from Artificial Analysis, and serving prices with the 262K context figure from OpenRouter. Release framing cross-checked against CNBC.
Frequently asked questions
What hardware do I need to run Nemotron 3.5 Lightning?
NVIDIA lists Blackwell (DGX Spark and GB10, GB200, GeForce RTX 5090), Hopper (H100 and H200), and Ampere through the W4A16 path. The single-GPU deployment targets it actually names are one H100 or one DGX Spark, so the consumer story is thinner than the RTX PC framing suggests. No VRAM floor is published for the 5090, which is the number we would want before planning a desk deployment. Checkpoints come in NVFP4 and BF16.
Is it really open weights, and under what licence?
Yes, the OpenMDW License Agreement version 1.1, which is permissive and has no user-count threshold bolted on. What makes this release stand out is that the weights aren't the only thing published: NVIDIA also released Nemotron-RL-Agentic-Terminal-Pivot, the agentic reinforcement learning dataset used to post-train the model for coding agent work. Weights are on Hugging Face, ModelScope, and build.nvidia.com as a NIM microservice.
How much does it cost through an API?
OpenRouter lists 0.05 dollars per million input tokens and 0.20 dollars per million output tokens. Artificial Analysis puts the blended figure at 0.07 dollars per million using their 7:2:1 cache to input to output ratio. There's also a free NVIDIA endpoint, and the terms on that one say prompts get logged for security and service improvement, so keep confidential material off it.
Does it really have a 1M token context?
The model card says up to 1M tokens. OpenRouter serves 262K. Both statements are published and nothing on either page reconciles them, so treat 262K as what you get today through the easiest route and test the long end yourself before designing around it. Artificial Analysis measured 49.19 on AA-LCR for the NVFP4 checkpoint, which is a long-context reasoning benchmark and the only public signal we have on how the model behaves out there.
What is NeMo Switchyard?
An open-source routing library that sits in front of a mixed pool of models and sends each request to whichever one suits it, without you rewriting the application. The idea is sound and the code is on GitHub. The headline number attached to it is not independent: NVIDIA says internal benchmarks show it holds frontier-level accuracy while cutting task completion cost to nearly one third of Opus 4.8 alone, which is NVIDIA measuring its own router against a competitor's model.
