What are NVIDIA's Nemotron 3.5 models, and which one fits your hardware?

NVIDIA is best known for the chips that train AI, but it also makes models. Nemotron is its family of open models, built for agents, programs that use a model to plan and call tools over many steps. Generation 3.5 arrived in stages: Lightning in August, Super VL and Bolt 74B-A9B on October 15, and Bolt E4B on October 23.

Why would a chip company give models away? These models are tuned for its own hardware, and its post makes clear that it wants organizations running agents on NVIDIA systems. That's a fair thing to keep in mind when you read its performance claims.

The family at a glance

A parameter is one learned number, and a token is a small piece of text. In a mixture-of-experts model, only some parameters work on each token, so "active" drives speed and "total" drives memory. Lightning, Super VL and Bolt 74B combine that with Mamba layers, a design that moves through a sequence without comparing every word with every other, along with some attention layers.

NVFP4 and GGUF describe formats. NVFP4 is a compact 4-bit format for NVIDIA chips, and GGUF is the file type read by llama.cpp, a popular engine for ordinary computers. Lightning and Super VL also support context lengths of up to a million tokens, but the memory for a long context comes on top of the file size.

One detail you might not expect: Bolt E4B is not trained from scratch. Its card says NVIDIA post-trained it from Google's Gemma 4 E4B to improve tool use and following instructions. It is a useful reminder that open models build on each other, and that a licence chain can follow.

A worked example: matching a model to a machine

Here's how to reason it through, using each card's listed hardware.

Take these rows as starting points. The Lightning card still marks its GeForce RTX 5090 recipe as unpublished, even though the hardware has the needed tensor cores, so a gaming card may take some extra work. Memory use with a real agent also depends on how much context you keep open.

How good are they?

NVIDIA reports its numbers under a consistent harness, and that matters more than the scores themselves. It also publishes recipes to reproduce them. On its own tests, Lightning at NVFP4 scores 52.8 on SWE-bench Verified and 75.57 on GPQA Diamond, which measures graduate-level science. It scores 10.47 on Humanity's Last Exam and 23.46 on Terminal-Bench 2.1.

Those are modest figures against the very large models in this guide. Lightning has only 3 billion active parameters, so it is a helper for quick tasks and sub-agents rather than a replacement for a frontier model. NVIDIA also says Super VL is comparable in accuracy to its largest reasoning model, Ultra, at about one-fifth of the parameters; that comes from NVIDIA's post and deserves your own test.

Openness is the family's real distinction. NVIDIA describes Nemotron as having open weights, training data and recipes, and the card links its pre-training datasets. Few labs at this level go that far. For a researcher, or an organization that must explain where a model's knowledge comes from, that is worth more than a few benchmark points. Read the data documentation for your exact model before calling it fully open.

What does the OpenMDW licence let you do?

Almost everything. OpenMDW 1.1 is a short licence written for models. It grants the right to deal in the "Model Materials" without restriction, including under copyright, patent, database and trade-secret rights. The 3.5 cards name it as the governing licence. It has no revenue threshold, no attribution banner and no use policy built in, and it places no restrictions on outputs.

It asks for very little in return. If you redistribute, keep a copy of the agreement and the original notices. If you sue claiming the materials infringe a patent or copyright, your rights end, unless you are responding to a suit brought against you first. You are also responsible for clearing other people's rights that might apply.

That last clause matters most for models like Bolt E4B, where another maker's work sits inside. Its card explains that Google's Gemma 4 is governed by Apache 2.0, also a permissive licence, so there is no conflict, but you must carry those notices too. And check the exact card for your file: the earlier Nemotron 3 models use a different NVIDIA licence.