NVIDIA Nemotron 3.5 Super VL: What the "5x Faster" Claims Actually Measure

Bottom line: NVIDIA’s October 15, 2026 blog says Nemotron 3.5 Super VL is 5x faster than the next-best model in its class. A separate Skydance workload also cites a 5x figure, but it measures latency in a specific deployment, so the two numbers should not be combined.

Three pages published on October 15 used the figure 5x for Nemotron 3.5 Super VL. They are not the same 5x. NVIDIA’s blog calls Super VL the fastest open agentic model in its class, 5x faster than the next best. Unite.AI, citing NVIDIA’s technical blog, says inference cost for a Skydance vendor-credit workload will decrease by 40x and latency by 5x. SQ Magazine, citing NVIDIA, says that on a single GPU the model delivers up to 5x higher throughput for agentic tasks than other open models in its size category. Class speed, a customer workload’s latency, and single-GPU throughput are three comparisons. Stapling them into “Super VL is 5x faster” drops the noun.

We already had a 120 billion parameter open model with 12 billion active, under a different name, in the March Nemotron 3 Super note. We already had a small model whose headline count was not the memory figure, in EmbeddingGemma 2. We already had a price that split depending on which sentence you read, in the Vidu Q4 note. Super VL is the same kind of split, on a speed claim, on the day the weights showed up.

The blog’s 5x is a class claim

The consumer blog’s sentence is about rank inside a class: fastest open agentic model in its class, 5x faster than the next best. It does not name the next best in that sentence. It does not say whether “faster” is tokens per second, time to finish a scripted agent task, or time to first token. A 5x with no named runner-up and no named clock is a marketing comparison. It can still be true on a chart NVIDIA drew. It is not a number you can reproduce from the sentence alone.

The same blog says Super VL delivers accuracy on multimodal agentic tasks comparable with NVIDIA’s largest reasoning model at roughly one-fifth the parameter count. Unite later prints Ultra’s scores next to Super VL’s. That is a score table. It is not an identification of “the next best” in the 5x sentence.

SQ’s 5x is narrower, and it is still NVIDIA’s. Barry Elad’s piece, timestamped 1:45 p.m. EDT on October 15, says that on a single GPU, Super VL delivers up to 5x higher throughput for agentic tasks compared with other open models in its size category, according to NVIDIA. “Up to” is a ceiling. “Its size category” is a bucket SQ does not define in that sentence. A single-GPU throughput ceiling is not the blog’s “5x faster than the next best,” even if both numbers came out of the same lab on the same day. One can be a class ranking. The other is a hardware-shaped throughput claim. Quoting either as “5x on one GPU versus the next best model” has merged them.

Skydance’s 5x is a workload, and one write-up drops it

Unite says Scale AI post-trained Super VL for Skydance’s vendor-credit finance agent and for EY’s M&A due diligence agent. The technical blog, as Unite cites it, states that inference cost for the Skydance workload will decrease by 40x and latency by 5x. “Will decrease” is a projection attached to a post-trained workload, not a measurement of the base checkpoint on a public leaderboard. The 5x here is latency on that workload. The 40x is cost on that workload. They are a pair. Using the 5x and leaving the 40x behind, or the reverse, edits Unite’s sentence.

SQ also covers Skydance. It does not, in that paragraph, print a latency cut of 5x. SQ’s closer says every speed and accuracy figure in the piece comes from NVIDIA. A 40x that appears in both Unite and SQ is the figure they share. If you read only SQ, the Skydance story is a cost multiple. If you read Unite, it is cost and latency. The technical blog URL both of them point toward is NVIDIA’s developer post. Unite’s attribution is the attribution.

Alice is the one with a percent: SQ says AI21 post-trained Super VL for Alice’s threat work and improved accuracy by 40%. The 40% accuracy and the 40x cost are both “40.” They are not the same unit. Accuracy points, or a relative accuracy gain, against a cost multiple, against a latency multiple. Three 40-shaped claims would be a mess.

120 billion is also what the March model said

Super VL is a 120 billion parameter mixture-of-experts model that activates 12 billion parameters per token. Unite prints both numbers. SQ prints both numbers. The active count is Unite and SQ, attributed to NVIDIA.

That pair is familiar on purpose, and it is a trap. In March, NVIDIA’s Nemotron 3 Super was also described as 120 billion total and 12 billion active. Same headline arithmetic. Different name: 3 Super then, 3.5 Super VL now. The March note is about that earlier checkpoint. This one is about a multimodal agent model released on October 15, with a vision claim the March headline did not carry. If a dashboard still labeled “Nemotron 3 Super” starts returning Super VL prices, or the reverse, the 120/12 line will not catch it. The name will. The VL suffix is the part that says the October model is being sold as text plus image, document, and video understanding. SQ says NVIDIA’s pitch is that teams can handle text and visual tasks without a separate vision-language model. That pitch is not a proof that the March Super checkpoint gained eyes. It is a claim about this release.

Unite says Super VL was distilled from Nemotron 3 Ultra, and that the technical blog credits Ultra’s coding and a SkewReverse loss for coding accuracy. Distilled from Ultra is not “is Ultra.” The one-fifth parameter remark on the consumer blog can sit next to a distillation story without becoming the same sentence. Ultra can be the teacher. Super VL can be the smaller student. The 5x class claim can still be about a different opponent than Ultra.

The score table is NVIDIA’s, and the clocks beside it are not 5x

Unite prints NVIDIA’s reported scores: 63.44% on Hermes Bench against 56.33% for Nemotron 3 Ultra, and 87.05% on LangChain behavioral evaluations against 84.3% for Ultra. On CharXiv, a chart-reasoning benchmark, 80.8% for Super VL against 63% for Nemotron 3 Nano Omni, rising to 84% when the model uses Python code interpretation. SQ prints the same Hermes, LangChain, and CharXiv figures, and its closer says the speed and accuracy figures come from NVIDIA. Two write-ups, one lab’s table. Agreement here is not independent replication. SQ also says an independent evaluation by EXO Labs, cited via the technical blog, placed the model on a Pareto frontier for accuracy and task-completion time. “Cited via the technical blog” is not EXO’s own paper in my hands.

CharXiv’s 80.8 and 84 are a pair the way the Skydance 40x and 5x are a pair. 80.8 is the score without the Python interpreter clause. 84 is the score with code interpretation, in Unite’s sentence. A chart that prints 84% and drops “when the model uses Python” has picked the assisted number. Nano Omni at 63% is the comparison Unite and SQ both name for CharXiv. Ultra is the comparison they name for Hermes and LangChain. Different baselines, different tests. A single “beats Ultra” line does not cover CharXiv, because CharXiv’s named opponent in these pieces is Nano Omni, not Ultra.

The other speed numbers should stay out of the 5x pile.

LiLiCorr, an open speculative-decoding method, delivers up to 2.3x faster token generation than decoding without speculation, Unite says. That 2.3x is against the same model without the trick, not against “the next best” open model. Speculative decoding speedups are a serving technique. They are not the class ranking.

SQ says Super VL delivers over 600 tokens per second on Artificial Analysis Intelligence Index agentic tasks on the Baseten endpoint, and that this is NVIDIA’s chart. 600 tokens per second is a rate on one endpoint. It is not a multiple. You cannot divide it into the 5x unless someone printed the other rate. Nobody in these three pieces printed the other rate next to the 600.

Unite’s Dynamo sentence is a serving configuration: on 16 B200 GPUs, at 50 tokens per second per user, Dynamo with KV-cache-aware routing serves up to 8x as many users while reducing first-token latency by 85%. The 8x is users at a fixed per-user rate, on 16 specific GPUs. The 85% is first-token latency in that setup. Neither is the blog’s 5x, and neither is Skydance’s latency 5x. A capacity planner who writes “5x to 8x faster, latency down 85%” has built a range the sources did not offer.

Unite says the recipe uses 4-bit weights and activations on Blackwell, with quantization-aware distillation meant to close the accuracy gap quantization introduces. The sentence does not give one.

Bolt’s 4GB date is October 23

The 24GB Bolt model is available now, in both Unite and SQ. Unite’s full name for it is Nemotron 3.5 Bolt 74B-A9B 24GB. It uses native 2-bit quantization to fit in 24GB and run locally on NVIDIA RTX and DGX Spark, Unite says, with llama.cpp, Ollama, LM Studio, Unsloth, and vLLM named as support. SQ calls it the 24GB Bolt model and prints the same software list. NVIDIA calls it the top open model in its class on PinchBench and the second-ranked open model on BrowseComp, according to both write-ups. “In its class” has the same hole as Super VL’s 5x: the class is not defined in the sentence.

The 4GB model is not in that “available now” clause. Unite schedules Nemotron 3.5 Bolt E4B 4GB for October 23, 2026. SQ says the 4GB Bolt model is coming October 23. Both agree on the date. Both say it is for smaller machines: GeForce RTX, Jetson, and RTX or DGX Spark, in the sentences they printed. A local-agent guide that tells you to pull the 4GB checkpoint on October 19 is early. The 24GB weights are the ones these pages say are public.

Training scale is Unite’s, and it is easy to confuse with the 2.8 trillion token dataset. Unite says Super VL was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples. Continued pretraining added 10 billion Indic tokens. Separately, NVIDIA is releasing a dataset of about 2.8 trillion tokens across 22 Indian languages, developed with IIT Madras and AI4Bharat. 26 trillion is a training-token claim. 2.8 trillion is a released dataset. 10 billion is a slice named inside continued pretraining. A model card screenshot that shows one of them is not a screenshot of the others.

The useful split, if you are about to repeat a number from this launch, is short. The blog’s 5x is a class ranking against an unnamed next best. SQ’s 5x is single-GPU throughput, up to, against other open models in an unnamed size category. Skydance’s 5x, in Unite’s citation of the technical blog, is latency on one post-trained workload, paired with a 40x cost line that SQ also prints and a latency line SQ’s paragraph does not. The March 120 billion / 12 billion model is not this checkpoint. The 4GB Bolt file is dated October 23. None of those sentences requires the others to be false. They require the noun.

What this means for you

When you see a speed multiplier in a launch post, check what it compares against and under which hardware. For your own evaluation, measure throughput and latency separately on the GPU you will actually deploy.