Nemotron 3.5 Super VL Matches Inkling Small at Less Than Half the Weights

Two models sit on 26 on the Artificial Analysis Intelligence Index this week, and the interesting part is what each one spent to get there. NVIDIA Nemotron 3.5 Super VL is a 120-billion-parameter mixture of experts with 12 billion active. Inkling Small, Thinking Machines' July release, is a 266-billion-parameter mixture of experts with 12 billion active. Same score, same active budget, less than half the weights on disk. Benchmark tables say what a model scores; the incident record says how releases like these fail in production, and those cases are catalogued in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That is the shape of last week in US open weights, and it is a better story than the launch post NVIDIA published on Thursday 15 October. The announcement paired Nemotron 3.5 Super VL with Nemotron 3.5 Bolt 74B-A9B 24GB — a 74-billion-parameter model squeezed into 24 GB of memory — and promised a third, Nemotron 3.5 Bolt E4B 4GB, for 23 October. Three models in eight days at three memory footprints is not a launch cadence; it is a portfolio argument, and the argument is that capability per gigabyte is now the number that matters.

What shipped, and on which day

The dates are worth stating precisely, because NVIDIA's own artefacts disagree by three days and the disagreement is informative rather than scandalous.

• Weights — the BF16 repository nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 was created on Hugging Face at 2026-10-12T15:32:06Z, with an NVFP4 sibling 24 seconds later and a pre-post-training base checkpoint 17 seconds before that.

• Announcement — the NVIDIA newsroom post carries a publication date of 15 October 2026. The model card's own release field reads 10/15/2026 at version 1.0 GA.

• Bolt 74B — its 24 GB repository predates the announcement considerably, created 7 October, with a GGUF build appearing on the Hub that evening. The blog presents it as shipping alongside Super VL on the 15th.

• Bolt E4B — announced for 23 October, no weights yet at the time of writing.

So the public artefacts arrived before the public prose. Anyone dating a deployment from the press release will be three days late to a model that was already downloadable, and anyone dating it from the repository will be quoting a timestamp NVIDIA has never acknowledged. Both are defensible; writing "released 15 October" without noting the 12th is the version that loses information.

The score is the boring part. The denominator is the story.

NVIDIA's pitch for Super VL is speed — "up to 5x faster throughput compared to other models in its class," attributed to the Artificial Analysis leaderboard. The claim is still partly unverified, and the section below deals with it. The durable fact about this release is a different one: the model scores 26 on the same panel where its March predecessor, NVIDIA Nemotron 3 Super 120B A12B, managed 13, and it does so at effectively identical size.

• Parameters — 120B total, 12B active, 512 routed experts per MoE layer. The March model was 120.6B total and 12.7B active. This is not a bigger model.

• Intelligence Index — 26 against 13, per Artificial Analysis, both flagged non-estimated. Roughly double, on the same parameter budget.

• Output tokens per Index task — about 49,000 against about 114,000. The new model finishes the same evaluation with roughly 57% fewer tokens, which is the mechanism behind a large part of the jump: it stops over-thinking.

• Against Inkling Small — both 26. Super VL does it with 120B total to Inkling Small's 266B, and both activate about 12B per token.

• Against Nemotron 3 Ultra 550B A55B — 23. NVIDIA's new mid-size model beats its own 550-billion-parameter flagship by three points while being roughly a quarter of the size.

The two efficiency axes get conflated constantly, so it is worth separating them. Active parameters govern what a token costs to generate. Total parameters govern what the checkpoint costs to store, load and keep resident. Super VL improves on both: it holds active parameters flat at 12B while roughly halving total parameters relative to Inkling Small, and it holds total parameters flat at 120B while roughly doubling its score relative to its own predecessor. A model that improved on only one axis would still be notable. This one did both in the same release.

Bolt is the quieter half of the release, and it has no independent score

The 24 GB model is the one most teams could actually run on hardware they own. Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter mixture of experts with 9 billion active, quantised to roughly 2.06 bits per expert during training rather than after it, a 262,144-token context, and the same OpenMDW-1.1 terms as its larger sibling. It is a Mamba-2 and Transformer hybrid with a C-RADIO v4-H vision encoder, so it accepts text, images and video and returns text.

NVIDIA's card reports, on its own harness: GPQA-Diamond 86.43, HLE without tools 25.30, Terminal-Bench 2.1 50.98, SciCode 41.02, BrowseComp 66.75, PinchBench 87.82 and WebArena vision 50.86. Every one of those is vendor-run and unreproduced. On the independent panel the model sits at 21 — a genuinely strong result for something that fits on a single 24 GB card, and also a reminder to be careful with the vendor table, because 86.43 on GPQA-Diamond next to a 21 overall index invites over-reading of a single row.

The download distribution tells its own story: the GGUF build of Bolt has drawn more than 87,000 downloads, against roughly 18,700 for the BF16 Super VL repository. The model people can run on a laptop is the model people are downloading. That is not a comment on quality; it is a comment on what "open weights" is actually for.

Where the US actually sits

The efficiency gain is real, and the leadership claim needs a qualifier NVIDIA's post does not supply. Among US-lab open-weight models on the independent panel, Meta Muse Spark Open scores 39 and is ahead of everything NVIDIA shipped last week. It is also gated on the Hub under a bespoke commercial licence, which puts it in a different adoption category from a checkpoint you can download under OpenMDW-1.1 without asking anyone.

• Muse Spark Open — 39, gated, bespoke licence.

• Nemotron 3.5 Super VL — 26, ungated, OpenMDW-1.1, 120B total / 12B active.

• Inkling Small — 26, permissive, 266B total / 12B active.

• Nemotron 3.5 Bolt 74B-A9B — 21, ungated, 24 GB footprint.

• GPT-5.5 Instant — 26, closed.

The Chinese open-weight field is still a different conversation. MiMo-V2.6-Pro sits at 46, GLM-5.3 at 45 and Kimi K3 at 44 on the same panel, so the strongest US permissively licensed model is roughly 20 points behind the strongest Chinese one. NVIDIA's answer to that gap is not a bigger model — it is post-training on your own data, with the datasets, recipes and reinforcement-learning environments published alongside the weights. Whether that closes anything is a question this release does not settle, and anyone claiming otherwise is reading intent into a benchmark.

What is measured, what is claimed, and what moved this week

At launch the 5x throughput claim had nothing behind it: Artificial Analysis listed the model with no speed figure at all. That has changed. The panel now records output speed at roughly 430 tokens per second, first of 28 in its comparison set. So part of the claim now has a second party: the model is fast, and not marginally so. What remains unverified is the multiple. "5x faster than the next best" needs a named next best and a matching harness, and neither the blog post nor the model card supplies one.

The price picture moved too, and it moved in a way that matters more than the headline rate.

• Median across providers — $0.225 per million input tokens and $0.575 per million output, with a 78% cache discount, per Artificial Analysis.

• Spread among providers that publish a rate — input from $0.10 to $0.20, output from $0.40 to $0.50 for the same checkpoint.

• Cost per Index task — $0.19, which is where the low price and the high token count meet.

• Verbosity — 170M output tokens consumed across the Index run against a median of 100M, flagged as somewhat verbose. Cheap output plus long output is not automatically cheap.

A 2x spread on input price for identical weights is the practical finding of the week. The model is the same file everywhere; the meter is not. That is an argument for treating provider choice as a routing decision rather than a procurement one — which is the piece of this story our own product happens to address, and the honest version of that claim is narrow. Nemotron 3.5 Super VL is not on our catalogue: the OrcaRouter model list returns a little over 200 model IDs and no NVIDIA entry today. What a router is worth here is that the moment a checkpoint does land, it lands at provider list price with 0% markup, and a model this new can sit behind automatic failover rather than at the root of a production path. Neither of those is a reason to buy the model. They are reasons not to sign a contract to try it.

Three things to watch, and one of them is nine days out

The first is Bolt E4B on 23 October. A 4 GB vision-language model that scores anywhere near its 24 GB sibling would be the more consequential release of the three, because it changes what hardware counts as enough.

The second is an independent evaluation of the quantised checkpoints. Every number in this article that comes from NVIDIA was measured on a full-precision or vendor-harness build. The NVFP4 file is the one most people will actually serve, and it has no published third-party score. NVIDIA's own claim is 99.62% median accuracy recovery against BF16 — a vendor figure, on a vendor harness, and the single most load-bearing unverified number in the release.

The third is whether the price spread narrows. A 2x gap between the cheapest and dearest server for one open checkpoint is a market that has not settled, and the direction it settles in will say more about the next six months of open-weight inference than any single benchmark row.

The score is the headline. The half-size checkpoint with the same active budget is the thing that will still matter in a year.