NVIDIA Nemotron 3.5 Super VL Is the Top Permissively Licensed US Open-Weight Model, and Its Weights Beat Its Own Announcement by Three Days

The timestamp is the part nobody has written down. The BF16 repository for NVIDIA Nemotron 3.5 Super VL — full name nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 — was created on Hugging Face at 15:32:06 UTC on 12 October 2026, alongside an NVFP4 build twenty-four seconds later and a base checkpoint nineteen seconds before that. NVIDIA's blog post announcing the model is dated 15 October. So for roughly seventy-two hours, the weights of NVIDIA's strongest open vision-language model were downloadable, indexed, and being quantized by third parties while the company that trained it had published nothing at all. That is not a leak and it is not a scandal. It is a release pipeline where the artifact ships before the prose, and it is worth stating plainly because the rest of this piece is about numbers, and the numbers have their own timestamps. On the Artificial Analysis Intelligence Index, NVIDIA Nemotron 3.5 Super VL scores 26.49, non-estimated. Its predecessor, NVIDIA Nemotron 3 Super 120B A12B, released 11 March 2026, scores 12.83 on the same panel. That is 2.06 times the score, and it arrives in 49,174 output tokens per task instead of 113,823.

Two things about that comparison deserve setting up before the detail. First, the 26.49 is a third-party number, measured on somebody else's harness with a published token budget, and it is flagged on the panel as not estimated. That matters more than usual here, because NVIDIA's own benchmark table is presented side by side with seven competitors and the two sets of figures do not always line up. Second, the leader claim in the headline is deliberately narrow. Among open-weight models built in the United States, Meta Muse Spark Open scores higher — 38.87, also non-estimated — but it ships under a bespoke commercial licence behind a gated repository, which puts it in a different adoption category from a checkpoint you can download and deploy under OpenMDW-1.1 today. Among permissively licensed US open-weight models on that panel, NVIDIA Nemotron 3.5 Super VL is first, and the next entry down is Inkling Small at 25.66.

What the repository establishes, and what only the panel does

The model card is unusually complete for something that shipped three days before its own press release, and it answers the structural questions cleanly. NVIDIA Nemotron 3.5 Super VL is a Mamba-2, mixture-of-experts and attention hybrid with a C-RADIO vision encoder and a vision-language connector, 512 routed experts per MoE layer, 120B total parameters with 12B active per token, and a context window up to 1M tokens. Native Multi-Token Prediction gives it speculative decoding without a separate drafter download. Inputs are text, images (JPEG, PNG) and video (MP4 up to two minutes, sampled as frames); output is text only. The licence is OpenMDW-1.1, the card states the model is ready for commercial or non-commercial use, and NVIDIA reports training across roughly 26 trillion tokens with about 3.8 million post-training samples and reinforcement learning across dozens of environments.

• Total / active parameters — 120B / 12B against the March model's 120.6B / 12.7B. Nearly identical, which is the point: the gain is not a bigger model.

• Intelligence Index — 26.49 against 12.83, both non-estimated, both per Artificial Analysis.

• Output tokens per index task — 49,174 against 113,823. The new model finishes the same task with 56.8% fewer tokens.

• Context and outputs — 1M tokens, text out only, on both. Neither is an image generator.

• Deployment floor — 4×H100 or 2×GB200 for BF16 on NVIDIA Nemotron 3.5 Super VL; an NVFP4 build exists for smaller footprints.

• Public traction — 142 downloads and 32 likes on the BF16 repository at the time of writing, against more than a million downloads on the March predecessor. The gap is three days of age, not three days of demand.

The capability breakdown on the panel is where the 26.49 stops being a single number. Super VL scores 38.47 in economics, 30.11 in engineering, 28.11 in legal and 27.19 in finance and accounting, but only 22.57 in strategy and operations. Its omniscience index is negative 4.22 — the model answers more of the panel's knowledge probes wrong than right, with a 23.3% accuracy rate against a 35.8% hallucination rate. That is not a defect unique to this release; it is what an unhedged open model looks like when you measure whether it declines to answer. On GDPval-AA, which grades finished professional deliverables rather than answers, it scores 1,162.41, and on the AA-Briefcase-Lite rubric it sits at an Elo of 998.76. Those two rows are the ones to watch over the next month, because they measure work rather than knowledge and they are the rows a 120B open model can plausibly keep improving through post-training.

The five-times claim, and why token counts are not seconds

NVIDIA's blog leans on speed: "5x faster than the next best," and the model "achieves up to 5x faster throughput compared to other models in its class." That is a vendor claim and nobody outside NVIDIA has reproduced it. It is also, on the current evidence, unsatisfiable — the same third-party panel that scores the model publishes no price and no latency figure for it, so there is nothing to check the claim against. What can be checked is a different quantity that often gets reported as if it were the same thing. Output tokens per task is a measure of how much the model talks, not how fast the server moves. 49,174 against 113,823 is real and it is a genuine efficiency gain — reasoning-heavy models burn their budget generating — but a model that emits 57% fewer tokens is not thereby 57% faster, and it is certainly not 5× faster. It is 57% fewer tokens. Hold those apart.

The vendor benchmark table has a second, smaller problem that a careful reader will notice in about ten seconds. Its first column is labelled "Nemotron 3.5 SuperGA BF16," not "Super VL." Given that the card's Model Version section says v1.0 — GA, that is most likely a release-candidate label that survived into the public document rather than a different model. Either way it is the kind of wart that tells you the documentation pass happened after the weights went up, which is consistent with the three-day gap. Against that column, NVIDIA reports GPQA-Diamond 89.52, HLE text-only 31.79, Terminal-Bench 2.1 66.70, SWE-bench Verified 71.10, BrowseComp 78 and Video-MME v2 32.85, with the full comparison column set running to Qwen3.5-122B-A10B, Qwen3.8-Flash-Next, DeepSeek-V4-Flash-0731, Nemotron 3 Ultra, Inkling-Small and GLM-5.3-Flash. All of it is vendor-reported through NeMo Gym, and NVIDIA publishes the recipes, prompts, parsers and scoring settings to let anyone rerun it. That is better practice than most, and it is still not an independent result.

Where it actually sits among US open weights

Sort the panel by open-weight models from US labs and the shape of the field is clear. Meta Muse Spark Open leads at 38.87 but is a 816B-parameter model under the Muse Model License, gated on the Hub, and a different proposition to adopt. Below NVIDIA Nemotron 3.5 Super VL at 26.49 sit Inkling Small at 25.66 and Inkling at 24.98 from Thinking Machines, both permissive. Then NVIDIA's own NVIDIA Nemotron 3 Ultra 550B A55B at 22.93, and NVIDIA Nemotron 3.5 Bolt 74B-A9B — released the same day as Super VL — at 21.41 for a 74B model that fits in 24GB of memory. The interesting structural fact is that Super VL beats its own company's 550B flagship by 3.6 index points while being less than a quarter of the size on disk and active-parameter count. That is the outcome the whole hybrid-MoE programme is aimed at, and on this panel it landed.

It also sets the real bar for the next few months. The gap to Muse Spark Open is 12.4 points, and the gap to the leading Chinese open-weight models is larger still — MiMo-V2.6-Pro at 46.32, GLM-5.3 at 44.78, Kimi K3 at 43.59. NVIDIA's answer to that gap is not a bigger model; it is post-training, on your data, on your racks, with the datasets and NeMo Gym environments published alongside the weights. Whether that closes the distance is an open question and this release does not settle it.

What it costs to run, and the one number that is missing

The BF16 checkpoint serves on 4×H100 or 2×GB200, with NVIDIA publishing vLLM and SGLang recipes and pinning specific container tags — vLLM v0.32.0 or a named nightly, SGLang v0.5.20 — which is a useful signal about how recently the kernels landed. The NVFP4 build is the one most teams will actually run. What nobody has published is a price. There is no per-token rate on the panel, no NIM list price in the announcement, and no third-party serving rate that a buyer could compare. For a model in this position the practical route is the one we run on: OrcaRouter puts the vendor's list price through with 0% markup across 200-plus models on one key, so when NVIDIA or a provider publishes a rate for this checkpoint it appears at cost rather than at a spread, and automatic failover means a day-old checkpoint can carry a slice of traffic while a proven model holds the rest. To be exact about what that does not mean today: this specific model is not on our catalogue yet, three days after its weights went up, because nobody is reliably serving it yet. That is a statement about availability, not about the model.

What to watch next

Three things will decide whether this release reads in six months as a step change or a footnote. The first is a second independent evaluation — the panel score is one, and one is not a trend. The second is whether the 5× throughput claim survives contact with a neutral benchmarker, because the token-count evidence supports efficiency and says nothing about latency. The third is the post-training ecosystem: NVIDIA lists Scale AI, AfterQuery, AI21, Mercor, Atlassian, CrowdStrike, Dassault Systèmes, Salesforce and others building on it, and the interesting question about an open model is rarely what the vendor shipped — it is what everyone else does with it afterwards. On the evidence available as of 16 October 2026, NVIDIA Nemotron 3.5 Super VL is the strongest permissively licensed open-weight model to come out of a US lab, it got there at constant parameter count by cutting its own reasoning budget by more than half, and it did so quietly enough that the first people to notice were the quantizers.