Nemotron 3.5 Super VL vs Nemotron 3 Ultra: Cheaper Than the Flagship, or Just Smaller?

NVIDIA has three Nemotron 3.5 tiers and one of them is a trap for planners. NVIDIA Nemotron 3 Ultra 550B-A55B, released on 4 June 2026, is the family flagship: 550 billion total parameters with 55 billion active per token, LatentMoE layers with Multi-Token Prediction, 1M tokens of context, and a stated minimum of sixteen H100s or eight GB200s. NVIDIA Nemotron 3.5 Super VL, whose weights arrived on Hugging Face on 12 October 2026 with NVIDIA's announcement following on 15 October, is the tier below it: 120B total, 12B active, the same 1M-token context, a C-RADIO vision encoder, and a floor of four H100s or two GB200s. NVIDIA's own card for Super VL benchmarks it directly against Ultra and claims accuracy in Ultra's neighbourhood at roughly one-fifth the parameters. That claim is vendor-reported, unreproduced, and it is the single sentence this whole comparison hangs on.

The arithmetic that actually decides deployment

Start with what the two checkpoints physically require, because that is the number a budget owner signs off on.

• Total parameters — 120B for Super VL against 550B for Ultra. Roughly 4.6 times the weight footprint, and the weights are what fill the GPUs.

• Active parameters per token — 12B against 55B. Compute per token scales with this, not with the total, which is why both models are far cheaper to run than their size suggests — and why the gap between them is 4.6× on memory but also 4.6× on per-token arithmetic.

• Minimum hardware — 4×H100 or 2×GB200 for Super VL; 16×H100 or 8×GB200 for Ultra. One node against a rack, in practical terms.

• Modalities — text, images and video in for Super VL; text in, text out for Ultra. Ultra has no vision path at all.

• Attention design — Super VL is the interleaved Mamba-2, mixture-of-experts and attention hybrid with 512 routed experts per MoE layer; Ultra uses NVIDIA's LatentMoE arrangement. Both carry MTP heads for speculative decoding.

• Licence — OpenMDW-1.1 with commercial use on both, weights and recipes open on both.

The pattern in those six lines is that Super VL is not a cut-down Ultra. It is a different shape of model: vision-capable, far smaller in memory, and — critically — deployable on hardware a mid-sized team might already own. Ultra's floor is a data-centre conversation. Super VL's floor is a serious inference node.

What the extra 43 billion active parameters buy

The general reasoning case

Ultra was built to sit at the top of NVIDIA's stack for pure text reasoning, and NVIDIA's June release benchmarks it against six contemporary frontier models across knowledge, math, coding and long-context suites. Those figures, like every figure in this comparison, come from NVIDIA's harness rather than an independent evaluator, so they establish what NVIDIA measured, not what a third party would.

The multimodal case

This is where the comparison inverts. Ultra cannot read a page. If the work involves documents, tables, charts, screenshots or video frames, the flagship is simply not a candidate, and the largest model in the family is irrelevant to your problem no matter how good its reasoning index is. Super VL's vision encoder and connector exist precisely for that workload, and its published strengths — OCR, structured extraction from visual documents, multistep business workflows — describe a different job from the one Ultra was tuned for.

The long-context case

Both reach 1M tokens. NVIDIA extended Super VL's context with a dedicated 35-billion-token stage followed by a long-context recovery pass, which is a stronger signal than a maximum length alone: it means long input was treated as a capability to preserve rather than a buffer to advertise. Ultra reached the same window through its own training pipeline. Without an independent long-context evaluation of both, treat this as a tie with a plausible edge to whichever model your own retrieval trace says holds up past 200k.

Serving one or both, and the tier trap

The trap is assuming the family is a ladder you climb. It is closer to a partition, and the partition is drawn by cost per task rather than by capability. A 12B-active model that needs two GB200s and finishes a document task five times faster than its class, if NVIDIA's throughput claim holds, beats a 55B-active model on that task regardless of which one is smarter — because the smarter one is idle waiting for the faster one's queue to drain. Flip the task to a long-horizon coding or math problem with no pixels in it and Ultra's extra parameters are doing real work.

Neither of these is on our catalogue, and it is worth being explicit rather than letting a vague sentence imply otherwise: OrcaRouter does not serve NVIDIA Nemotron 3.5 Super VL, Nemotron 3 Ultra 550B-A55B, or any other NVIDIA checkpoint. They are weights you deploy. What a routing layer contributes to a decision like this is measurement rather than access. The routing DSL lets you put the same prompt in front of two upstream models and read the answers side by side, so an evaluation of a 12B-active candidate against a 55B-active incumbent is a config change instead of two integration projects. Automatic failover matters for the opposite reason: if a hosted model you route to degrades mid-pipeline, the request moves rather than fails, which keeps a fallback path alive for the teams who are keeping Super VL self-hosted as their primary. And for the parts of the workload that no self-hosted tier should absorb — bursts, rare modalities, one-off large-context jobs — the catalogue behind one key, more than two hundred models at 0% markup over provider list price, is where those requests go.

The honest framing of that paragraph is that we are not the vendor of either model here. We are the layer that makes the comparison cheap to run and the escalation path cheap to keep.

The claim to test before you commit

Everything above reduces to one unverified sentence: that a 120B-parameter model with 12B active matches NVIDIA's largest reasoning model at roughly a fifth of the parameter count. If it holds, Super VL is the better default for a mixed workload and Ultra is reserved for the narrow tail of problems that genuinely need it. If it does not hold — and a day after release nobody outside NVIDIA has measured — then the two tiers are simply different sizes of the same thing, and the choice collapses back to what your task needs: pixels or no pixels, one node or one rack.

The test is cheap and specific. Take fifty real tasks from your own queue that have a checkable answer, run them against both, and score completion rather than benchmark scores. That number, not NVIDIA's, is the one that decides which tier you buy hardware for — and it is the number that will still be true when the next Nemotron release resets this page.