Chat with Nemotron 3.5 Lightning.

Wakes up just 3B of itself per word, so it answers fast for its size.

Start chatting with Nemotron 3.5 Lightning now — free, no account

At a glance

  • Maker: NVIDIA
  • Released: 2026
  • Parameters: 30B
  • Active per token: 3B
  • Context window: 262k
  • Size: Medium
  • Memory: Handles long documents
  • Good at: Fast, high-volume work
  • Supports: zero data retention
  • License: NVIDIA Open Model License
  • Upstream API price: $0.25 per million tokens

Announced in NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard.

What is Nemotron 3.5 Lightning?

Nemotron 3.5 Lightning is NVIDIA's speed-first open model, built for work where the answer needs to be good enough, now. NVIDIA is better known for the GPUs that train everyone else's models, but its Nemotron line has grown into a serious open-weight family in its own right, and Lightning is the tier where the company pushes latency as far as it will go: 30 billion parameters in total, of which only about 3 billion wake up for any given token.

That 3B-of-30B design is one of the most aggressive sparsity ratios on UsingOpen, and it is the whole point of the model. Fewer active parameters per token means faster time-to-first-word, higher throughput and lower cost — $0.25 per million tokens on NVIDIA's API — at the price of some depth on the hardest problems. The 262k-token context window, roughly 200,000 words, keeps it useful on long documents even as it stays quick.

Two honest limits shape where it fits: it does not take image input, and it does not support tool calling. This is a text-in, text-out engine. Within NVIDIA's own line it is the volume tier, with the much larger Nemotron 3 Ultra above it for reasoning and orchestration. If your job is transforming, classifying, extracting or summarising text at scale — thousands of calls, not three — Lightning is exactly the shape of model you want.

What can you do with Nemotron 3.5 Lightning?

Lightning's job description is high-volume text work. Good fits include:

  • Classifying, tagging and routing large batches of text — support tickets, articles, inbox triage.
  • Extracting structured data such as names, dates, amounts and entities from long documents.
  • Summarising reports and transcripts, with the 262k-token window — roughly 200,000 words — keeping whole documents in one call.
  • Rewriting and copy-editing passes where speed matters more than brilliance.
  • Draft-first workflows: generate many rough options fast, then let a bigger model refine the best one.
  • Anything with a strict latency budget, thanks to the 3B-active-per-token design.

How does Nemotron 3.5 Lightning compare?

Inside NVIDIA's family the choice is stark: Lightning is built for throughput, while Nemotron 3 Ultra — 550B parameters with 55B active, at $2.40 per million tokens — is built to reason and even to direct other models. They are less competitors than the two ends of a pipeline: Ultra decides, Lightning executes at scale.

Across makers, Lightning's rivals are the other cheap sparse models: DeepSeek's V4 Flash costs less and adds tool support with a million-token window (though no images either), and Google's Gemma 4 26B is a similarly frugal mixture-of-experts that also reads pictures. If your workload is pure, high-volume text, Lightning's latency is the argument; if you need tools or sight, look next door.

Is Nemotron 3.5 Lightning right for you?

Pick Lightning when throughput is the requirement and text is the material: classification pipelines, extraction jobs, bulk summarisation, rewriting passes across thousands of documents. At $0.25 per million tokens with a 3B-active design, it is built for exactly the work that would bankrupt a flagship model, and the 262k window keeps whole documents in single calls. It is also the model for latency-sensitive products — autocomplete-adjacent experiences, live suggestions, on-the-fly classification — where time to first token matters as much as the answer itself.

Do not pick it for work that needs sight, tools or deep reasoning: it has none of the three, by deliberate design. Agent loops belong to the tool-supporting models, images belong to the multimodal tier, and hard analysis belongs to the reasoners — including its own big sibling, Nemotron 3 Ultra. Lightning's genius is knowing what it is not. If your text volume is mostly code rather than prose, DeepSeek's V4 Flash brings tool support and a million-token window at a lower price.

A common architecture that works well: use a reasoner to design the pipeline and Lightning to run it. The Nemotron family is practically built for that split — Ultra decides, Lightning executes — and the two share a licence and a lineage that make them easy to deploy together.

Frequently asked questions

What is the context window of Nemotron 3.5 Lightning?

262k tokens — roughly 200,000 words. Long enough to keep whole reports, transcripts or code modules in a single call while the model stays fast. That is long enough for most single documents and generous for the batch work the model is built for — volume jobs rarely need archive-scale memory.

Does Nemotron 3.5 Lightning support image input?

No. Lightning is text-only by design — image understanding is not part of this tier, and the stripped-down input set is one of the reasons it stays so quick.

Does Nemotron 3.5 Lightning support tool calling?

No. It is a text-in, text-out engine without tool support, which makes it a poor fit for agentic loops but an excellent one for high-volume transformation, classification and summarisation. For agentic loops the answer is a sibling: Ultra at the centre of the pipeline, Lightning nowhere near it — the family’s roles are unusually well separated.

Who makes Nemotron 3.5 Lightning?

NVIDIA, the GPU maker, as part of its Nemotron family of open models. It was announced together with the NeMo Switchyard tooling on the NVIDIA blog. The NeMo Switchyard tooling that accompanied its release is aimed at exactly these deployments, which tells you who NVIDIA expects to use it.

Why is Nemotron 3.5 Lightning so fast and cheap?

Sparsity. Of its roughly 30 billion parameters, only about 3 billion are active for any given token — one of the leanest active ratios on UsingOpen. That keeps time-to-first-word and cost per million tokens ($0.25) very low. The ratio is the story: one parameter in ten awake. Nothing else on UsingOpen strips down further, which is why nothing else answers this quickly.

Is Nemotron 3.5 Lightning free to chat with on UsingOpen?

Yes. Pick it in the model picker and start typing — no account, no payment. History is saved only on your device, and the picker shows live availability. Benchmarking a throughput model for free is the sensible first step — paste a real batch through it before committing any budget.

Related models on UsingOpen

Also worth a look if you are weighing Nemotron 3.5 Lightning up:

  • Nemotron 3 Ultra by NVIDIA: the largest open NVIDIA model — 550B total, 55B awake, built to reason and orchestrate.
  • DeepSeek V4 Flash by DeepSeek: the bargain coder — 13B of 284B awake, a million-token window at $0.18 per million tokens.
  • Gemma 4 26B by Google: the mixture-of-experts Gemma — 25B total, 3.8B awake, quick on a budget.

Head-to-head comparisons

See how Nemotron 3.5 Lightning lines up with the models people most often weigh it against:

Model availability changes as upstream capacity comes and goes — the picker on the chat page shows what is ready right now.

Chat with Nemotron 3.5 Lightning free on UsingOpen — no account, no sign-up

← All open-weight models · Compare models · Free AI chat · FAQ