On openness · The But Is It Open? blog

Jensen Huang Called Nemotron Open Source. How Open Is It Really?

· 6 minute read

Jensen Huang championed open AI. We examine what NVIDIA's Nemotron 3.5 Lightning actually releases, where its training recipe falls short, and what his departure could mean.

Abstract painting on warm ivory paper, with rough ochre and translucent gray planes, a narrow burgundy stripe, and a dark navy block.

Jensen Huang Called Nemotron Open Source. How Open Is It Really?

Jensen Huang made a promise about NVIDIA's AI models that is worth checking now that he has left the company. In a conversation with Lex Fridman in March, he said of Nemotron 3: “We open sourced the models, we open sourced the weights, we open sourced the data, we open sourced how we created it.”2

That's a bigger claim than “you can download the weights.” NVIDIA has put real money and engineering into giving developers models, tools, datasets and recipes to work with. That deserves credit. The question is how much of the model you can carry away, and how much of its creation you can repeat.

I wrote last week about what Huang's departure could mean for Nemotron. This time I want to look at the model itself and at what he actually argued for. NVIDIA announced on November 12 that Huang had resigned as CEO and board member, effective immediately; CFO Colette Kress is interim CEO.1

What Huang meant by open

Huang's case was both public-minded and commercial. In the Fridman interview, he said proprietary models have a place, but that if everything is proprietary, research and building on the work become harder. He also said developing models helps NVIDIA understand what future computing systems need.2 More people running AI can mean more customers for NVIDIA's hardware. A business reason doesn't make the public benefit imaginary; it does help explain why the company invested.

In July he shared an industry letter, Open Weights and American AI Leadership. It argues that downloadable models give organizations more control over their data and reduce dependence on a single provider. It also acknowledges a difficult tradeoff: once weights are released, modified versions are hard to trace or recall. The signatories argue that wider access helps defenders and researchers, and that restrictions should target demonstrated harms rather than shut down open models altogether.3

That's an argument, not a settled safety result. Critics have a serious point too: a capable model that can be copied freely can also be adapted for harmful use, and a later safety fix can't simply be imposed on every copy. The letter recognizes the loss of control but weighs the benefits differently.

What you get with Lightning

Nemotron is a family, not one model with one license. To keep this fair, I'm looking at one specific release: the full-precision Nemotron 3.5 Lightning 30B-A3B, released in August. Its model card describes downloadable weights, training stages and ways to run the model.4 My assessment below is provisional and applies to this checkpoint, not every Nemotron release. The openness index's family overview deliberately leaves the mixed-license family unscored.

Access — strong. The Lightning weights are publicly downloadable without an approval gate. NVIDIA also provides model documentation, evaluation instructions and training resources.11 A download is not a token gesture: it lets someone keep and work with a model without a continuing API relationship.

Permission — broad, with caveats. The weights use OpenMDW 1.1, which grants broad rights to use, modify and share, including commercially. Distributors must retain the license and applicable notices. It also contains a termination clause concerning certain patent or copyright lawsuits, and puts third-party rights clearance on the user.6 OpenMDW is not on the OSI's approved-license list as of this writing; that is a status fact, not a claim that it prohibits commercial use.13 Other Nemotron checkpoints and datasets can have different terms.

Deployment — strong. You can run Lightning yourself using open serving software such as vLLM or SGLang instead of NVIDIA's hosted API. NVIDIA's full-precision example uses an 80 GB H100 or A100 GPU; that's independent hosting, not a promise that it fits on an ordinary laptop.4 NVIDIA also publishes a smaller-format variant and local deployment guidance.

Transparency — substantial, but incomplete. NVIDIA names training stages, describes data sources and filtering, publishes a technical report and offers large training datasets.410 That's more useful than a model card that tells you almost nothing. But the card also identifies private datasets, and says some code, math and multilingual data require approval. The information is much better than none; it is not a complete public copy of everything used.

Reproducibility — limited. This is where I would qualify Huang's broad claim. NVIDIA's own Lightning training guide says its public recipes use an open subset of the training data and will produce different results from the published benchmarks, which used additional proprietary data. The guide also says its pretraining recipe is a reference implementation, not a reproduction of the original distillation run.9 You can learn from the recipe and build on it. You should not assume it recreates the released checkpoint.

Governance and participation — a real path, within limits. NVIDIA's developer repository invites outside contributions to examples, recipes and tools, with published pull-request guidelines.512 That's worth recognizing. A contribution path to those tools isn't the same thing as shared authority over NVIDIA's decisions about future model training or releases.

On the book's release-specific rubric, my provisional scores are Access 3, Permission 2, Deployment 3, Transparency 2 and Reproducibility 1, out of three each. That makes this Tier 3: practical open-weight under the current gates. Governance and participation sit separately; I'd provisionally put them at 2, real but bounded. These are measurements of this release, not a grade for its usefulness or an official OSI certification.

You can take Lightning home. You cannot take home everything NVIDIA used to make it.

What happens without its champion?

AI researcher Nathan Lambert calls Huang's departure a major setback for the US open-model ecosystem. He expects funding Huang championed to dry up and worries that open models have lost an influential advocate against poorly targeted regulation. Lambert also acknowledges NVIDIA still has business reasons to support them.7 Those are informed predictions, not an announcement that funding has been cut.

NVIDIA's board says it expects its operations and product roadmap to continue without interruption.1 That is not a specific promise about future Nemotron datasets, licenses or research budgets. I would watch the next releases: are the weights still downloadable, are the recipes maintained, and do the data disclosures get better or thinner? The company's interest in selling hardware may support continuity, but it doesn't answer those questions by itself.

I don't expect every company to publish every piece of training data. Rights, privacy and cost are real constraints. I do want the public description to match what someone can actually obtain and rebuild. NVIDIA has made a useful, unusually documented release; the gaps matter precisely because the work is worth building on.

So, is Nemotron 3.5 Lightning open? Yes, in the practical sense that you can download, adapt and run it independently, with broad permissions. It is not an entirely reproducible open training system. Both parts of that answer matter.

Research based on NVIDIA's model card, training guide, license and statements, and Nathan Lambert's commentary, checked November 17, 2026.

← More from the blogJoin the newsletter ↗