NV-Reason-CT vs Nemotron 3 Ultra: One Logo, Two Release Philosophies

NVIDIA shipped two open-weights models this year and told nobody about one of them. Nemotron 3 Ultra landed on Hugging Face on June 4, 2026 as a 550B-total, 55B-active flagship with a technical report, an OpenMDW-1.1 licence regime, and an entire open pre-training and post-training dataset release behind it. NV-Reason-CT landed on September 8, 2026 as a 4.69B-parameter CT specialist with a repository, a demo Space, and an arXiv preprint two weeks later that mentions the release in a citation block. There was no announcement for the second one. Compare them and you are not really comparing capabilities — a 550B general flagship and a 4.69B medical specialist do not overlap. You are comparing two different theories of how to release a model, both operated by the same company, three months apart.

What each repository actually contains

The Nemotron 3 Ultra release is a complete open pipeline. The NVFP4 checkpoint on Hugging Face has drawn 249,746 downloads and 338 likes; the BF16 sibling, the base checkpoint, and the post-training datasets are all published alongside it. It is twelve-language, text-only, built on a latent mixture-of-experts Mamba2-Transformer hybrid with multi-token prediction, and it carries the openmdw-1.1 licence. Artificial Analysis places it at an Intelligence Index of 47.7 in the shipped NVFP4 precision and 48.2 in full precision — the highest of any US open-weights model, per that board. Our own earlier coverage of the Nemotron teacher models records the median provider price at roughly $0.60 per million input tokens and $2.60 per million output, and NVIDIA's 183-output-tokens-per-second throughput figure in third-party measurement.

NV-Reason-CT is a much smaller and much more enclosed artifact. Ten point six gigabytes of BF16 weights, a 10 KB model.py, a 26 KB processor.py, a chat template, and an inference.py. One demo Space. One paper. Two hundred and ten downloads as of this week. No dataset release. No technical report. No serving configuration beyond a single-volume example.

The contrast is not about size — it is about what a downstream engineer can do with the artifact. Ultra's release is designed so that someone else can reproduce the training. NV-Reason-CT's release is designed so that someone else can run the inference. Those are different promises.

Only one of them is built on NVIDIA's own stack

Here is the detail that surprises people, and it is verifiable from the model card's frontmatter: NV-Reason-CT's base model is Qwen/Qwen3.5-4B, with a fine-tune relation, not a Nemotron checkpoint. NVIDIA bolted a Primus 3D vision transformer — initialised from COLIPRI weights — onto a Qwen language model and trained the pair end-to-end with supervised fine-tuning followed by Group Relative Policy Optimization.

That is not a criticism and it is not a scandal. It is the industry norm for vision-language work, and NVIDIA discloses it plainly in the repository metadata. But it does mean the single most natural assumption about this matchup — that NV-Reason-CT is a Nemotron derivative — is false. The two models on this page share a logo and a licence family and essentially nothing else. No shared architecture, no shared tokenizer, no shared training data, no shared serving path.

Size, in the numbers that decide where it runs

• Parameters — NV-Reason-CT 4.69B total / 4.35B active in the CT pathway vs Nemotron 3 Ultra 550B total / 55B active sparse LatentMoE • Checkpoint on disk — NV-Reason-CT ~10.6 GB BF16 vs Nemotron 3 Ultra hundreds of gigabytes in BF16, published additionally in NVFP4 • Input — NV-Reason-CT 3D NIfTI CT volumes in Hounsfield units, chest or abdomen only, one 13,824-token volume prefix vs Nemotron 3 Ultra text, up to 1M tokens of context with a 65K output cap • Context — NV-Reason-CT not applicable in the usual sense vs Nemotron 3 Ultra 1,000,000 tokens • Licence — both OpenMDW-1.1 • Hardware — NV-Reason-CT Ampere / Hopper / Lovelace, tested on H100 and L40S vs Nemotron 3 Ultra multi-GPU serving at 55B active • Independent score — NV-Reason-CT none published vs Nemotron 3 Ultra AA Intelligence Index 47.7 NVFP4 / 48.2 full precision

The practical line is the memory one. A 4.69B model with a 3D encoder fits on a single card and a research group can justify that. A 550B-total model at 55B active needs real serving infrastructure even quantised. Neither is a weekend experiment, but they are on different orders of that problem.

Two evaluation regimes that cannot be averaged

Ultra's evidence is general capability: an Artificial Analysis Intelligence Index in the high forties, benchmark coverage across reasoning, coding and agentic tasks, and vendor-reported figures for the family including SWE-Bench Verified 71.9, MMLU-Pro 86.8 and LiveCodeBench v6 89.0 — those last three being NVIDIA's own numbers and labelled as such.

NV-Reason-CT's evidence is one table. Macro-F1 0.614 and Macro-AUROC 0.871 on CT-RATE's 18-label classification task, at a fixed uniform threshold, using a direct Yes/No prompt with no classification head, against VoxelFM at 0.581/0.870, Pillar-0 at 0.544/0.861, ClinFusion-8B at 0.442, CT-CLIP at 0.398/0.733, Merlin at 0.358/0.662 and MedGemma 1.5 at 0.303. NVIDIA's numbers again, on the model card, with no independent reproduction yet.

An AA Intelligence Index of 47.7 and a CT-RATE Macro-F1 of 0.614 are not two measurements of one thing. They are measurements of two things. There is no defensible way to say Ultra is "better" than NV-Reason-CT, and the reason is not scoring preference — it is that the instruments do not overlap at all. The only honest comparison in this pairing is of who each model is accredited for, in the literal sense of what evidence exists that it does its job.

Why the quiet release is the rational one

Put yourself on the medical-imaging team. You have a working 3D CT model. You want radiologists, medical students and researchers to use it. What does a keynote get you?

A launch event raises expectations about support, roadmaps and longevity that a research team cannot meet. It invites general-purpose evaluation on benchmarks where a 4.69B modality-locked model will look unimpressive for structurally irrelevant reasons. And it puts a model that explicitly "must not be used as a substitute for professional clinical judgment" in front of an audience that does not read licence terms. A repository, a paper and a demo Space reach exactly the audience that reads all three, and they let the artifact speak at its own speed. The version string is "0.1". The card says research and education only. That is a release designed to be cited, not bought.

Ultra's release is the opposite and is equally rational — a flagship needs distribution, a rate card and ecosystem support, which is why it came with datasets and a report.

Where a router fits, and where it cannot

Neither of these models is on OrcaRouter, and I am not going to suggest otherwise: NV-Reason-CT is a 10.6 GB checkpoint with custom model code you host yourself, and Nemotron 3 Ultra at 55B active is served through NVIDIA's own API and several third-party platforms.

What a routing layer does for a team working with both is narrower and still useful. A clinical research pipeline that uses NV-Reason-CT for the volume and a general model for the surrounding text needs a place to swap that general model without a second contract, and one OpenAI-compatible endpoint covering 200+ models with automatic failover across providers is that place. It keeps the self-hosted specialist and the hosted generalist under one key rather than two integrations.

What to take from the pairing

Read as competitors, the two models are a category error. Read as a case study, they are unusually clear. Same vendor, same licence family, same year — and two release strategies chosen to match two completely different downstream intents. One is infrastructure with an ecosystem attached. The other is a paper with weights attached, released so that a specific small audience can run it and cite it.

If you are looking for a general open-weights model, NV-Reason-CT is not a candidate and was never meant to be. If you read chest and abdominal CT volumes programmatically, Nemotron 3 Ultra cannot help you and was never meant to. The only real overlap between them is the logo, and the only real lesson is that NVIDIA is willing to ship quietly when the audience is small and technical enough that it will find the repository unaided.