NVIDIA Nemotron 3.5 Super VL Ships: 120B Open Weights, a 5x Claim, and a 249 GB Catch

NVIDIA Nemotron 3.5 Super VL is out, and the specification that should decide whether you touch it is not the one the announcement leads with. The model is a 120B-parameter vision-language system with 12B active parameters, a 1M-token context, native speculative decoding and support for text, images, documents and video. NVIDIA's launch post is dated 15 October 2026 and pitches it as the fastest open agentic model in its class — "up to 5x faster throughput compared to other models in its class," measured on the Artificial Analysis Intelligence Index leaderboard. What the post does not put in front of you is that the release is two checkpoints, that the difference between them is 174 GB of VRAM, and that the smaller one pays for that saving unevenly.

What actually shipped on 15 October

Two things happened on different days, and the gap matters if you are dating a deployment. The weights landed on Hugging Face on 12 October at 15:32 UTC. The announcement followed on 15 October. NVIDIA's own model card gives the release as 10/15/2026 at version 1.0 GA, so the vendor considers the 15th the ship date; the repository timestamps say otherwise. Neither date is more than a week old, and both are recent enough that nothing about this release has had time to be independently reproduced.

The release is broader than the model the announcement is named after:

• Nemotron 3.5 Super VL — the flagship, 120B total / 12B active, in a BF16 build and an NVFP4 build, both published 12 October.

• Nemotron 3.5 Super VL Base — the pre-post-training checkpoint, published alongside the instruct build, which matters because the licence lets you run your own post-training.

• Nemotron 3.5 Bolt 74B-A9B 24GB — an edge model that fits 24 GB through native 2-bit quantisation, published 7 October.

• Nemotron 3.5 Bolt E4B 4GB — the smallest of the family, dated 23 October.

The architecture is a Mamba-2/Transformer hybrid Mixture-of-Experts with Multi-Token Prediction, 512 routed experts per MoE layer, a C-RADIO vision encoder and a vision-language connector. Multimodal handling covers OCR, tables, charts, multi-page documents and visual grounding; video is accepted as MP4 up to two minutes, sampled at 2 FPS and up to 256 frames, or 1 FPS and 128 frames at 1080p. The licence is OpenMDW-1.1, which the model card states is ready for commercial or non-commercial use — a genuinely permissive term set, and a real part of why this release got attention.

Two checkpoints, and the 174 GB between them

This is where the release gets interesting for anyone who has to pay for GPUs. NVIDIA published the same model twice, at two precisions, and the two builds are not close in size. Both figures below are read from the Hugging Face repository file listings, which report the sizes of the individual weight shards:

• BF16 checkpoint — 248.94 GB across 63 shards. The safetensors header puts the parameter count at 124.45 billion.

• NVFP4 checkpoint — 75.06 GB across 88 shards. Inference runs W4A4 on Blackwell and W4A16 on anything older.

• Hardware, full precision — NVIDIA's card lists 4× H100 or 2× GB200 for text, image and video.

• Hardware, quantised — 1× H200, 2× H100, 1× GB200, or a single DGX Spark, with the Spark path needing a LiLiCorr drafter for speculative decoding.

• The saving — 173.88 GB, or about 3.3× smaller on disk, and the difference between a four-GPU node and a single accelerator.

Read that as an economics question, because that is what it is. Going from four H100s to one H200 is not a tuning decision; it is the difference between a model you can put behind a product this quarter and one you cannot. The quantisation recipe NVIDIA used is also more careful than the usual: the vision encoder, the connector, the Mamba convolutions and the embeddings stay in BF16, the MoE router stays in FP32, and the rest is produced by quantization-aware distillation from the BF16 model with 200 iterations of SFT-based distillation on top.

What the quantised build gives up

NVIDIA published a paired table comparing the two checkpoints with every other setting held static, which is the honest way to run this comparison and rarer than it should be. The pairs below are NVIDIA's own numbers, from the NVFP4 model card, and they are not all in the same direction.

Where the quantised build loses:

• Terminal-Bench 2.1 — 66.70 in BF16 against 60.80 in NVFP4. A 5.9-point drop, the largest single regression in the table.

• WebVoyager, offline — 77.72 against 66.31. An 11.4-point fall on browser-agent work.

• DeepSWE, offline — 24.65 against 18.20.

• MMLU-ProX — 76.91 against 73.76.

• tau3-bench banking — 18.88 against 17.78.

Where it gains, or holds:

• GPQA-Diamond — 89.52 against 89.90.

• APEX shortlist — 88.36 against 89.36.

• AA-Omniscience hallucination rate — 30.0 against 27.7, i.e. the quantised build hallucinates slightly less on NVIDIA's own measure.

• tau2-bench telecom — 73.19 both sides, identical.

• OCRBench v2 EN — 68.12 against 68.34.

• MMLongBench-Doc — 62.82 both sides.

The pattern is legible without a statistician. The document, OCR and knowledge-recall workloads barely move, and some of them improve. The long-horizon agentic workloads move a lot. If your application reads filings, receipts or page images and returns structured output, the 75 GB checkpoint is what you should be serving and you are giving up close to nothing. If your application is an agent that runs a terminal or drives a browser for dozens of consecutive steps, you are trading roughly nine percent of Terminal-Bench for three-quarters of your GPU bill, and that is a decision worth running on your own eval rather than on a vendor table.

The 5x, and what it is measured against

NVIDIA's headline is throughput, and it deserves to be read precisely rather than dismissed or repeated. The claim is "up to 5x faster throughput compared to other models in its class," and the class is not defined in the post. The measurement is NVIDIA's, on NVIDIA's serving stack, at whatever concurrency and sequence mix the recipe chose — the model card's own deployment tables specify 32 concurrent sequences and 5 or 3 speculative draft tokens depending on hardware, and throughput on a 20-turn agentic proxy is a very different number from throughput on single-shot prompts.

None of that makes the claim false. It makes it a vendor claim about a vendor-configured workload, which is exactly what the label says. What is worth noting is that on this release there is no counterweight: no independent throughput measurement of these checkpoints has been published, so the only numbers in circulation are NVIDIA's. A throughput figure without a named baseline model, a stated concurrency and a stated hardware configuration is not something you can plan capacity against. If the 5x is load-bearing for your decision, reproduce it on your own traffic before you sign a GPU reservation.

Where the model is strong, by NVIDIA's own table

The single-model card also lists how the BF16 checkpoint compares with its peers, and it is worth reading the columns to the right of NVIDIA's own, because they are also NVIDIA's. Against Qwen3.5-122B-A10B, Qwen3.8-Flash-Next, DeepSeek-V4-Flash, Nemotron 3 Ultra, Inkling-Small and GLM-5.3-Flash:

• Video-MME v2 — 32.85, against Qwen3.8-Flash-Next at 45.1 and GLM-5.3-Flash at 14.8. Middle of the field, and the widest spread of any row.

• OCRBench v2 EN — 68.12, roughly level with Qwen3.8-Flash-Next at 67.9 and Qwen3.5-122B-A10B at 68.3.

• CharXiv reasoning — 80.82, behind both Qwen models.

• MathVision — 78.16, behind Qwen3.8-Flash-Next at 87.5.

• MMLU-ProX — 76.91, behind every peer column on that row.

• AA-Omniscience hallucination rate — 30.0, against Qwen3.5-122B-A10B at 84.5 and DeepSeek-V4-Flash at 95.9. On this row, lower is better, and the margin is not small.

• Terminal-Bench 2.1 — 66.70, against Qwen3.8-Flash-Next at 85.8 and GLM-5.3-Flash at 82.7.

• SWE-bench Verified — 71.10, against GLM-5.3-Flash at 79.10.

• OSWorld-Verified — 51.14, versus GLM-5.3-Flash at 14.78.

The shape of that is consistent with what NVIDIA says the model is for. It is not the leader on raw reasoning or coding, and it does not claim to be — the announcement positions it against its own largest model at roughly one-fifth the parameter count, not against the frontier. It is strong on hallucination discipline, document extraction and desktop-agent work, mid-pack on video, and behind on the hardest agentic-terminal benchmarks. If you came to this release because "video understanding" was in the first line of the announcement, the video numbers are the ones to look at closely.

What it costs to run

There is no list price, and that is not an omission on NVIDIA's part — it is the shape of an open-weights release. NVIDIA does not sell this model per token at a published rate; it publishes weights under OpenMDW-1.1 and a NIM container with validated configurations, and you pay for the GPUs. So the cost question is a GPU-hours question, and the two acceptable answers are 4× H100 for BF16 or one H200 or a single DGX Spark for NVFP4.

That is the part of the arithmetic most teams get wrong when they compare an open model against a hosted API. The hosted side quotes a per-token rate that already contains someone else's utilisation curve. The open side quotes a checkpoint and a minimum hardware list, and the real number depends entirely on whether you can keep the GPU busy. A single H200 running a 75 GB checkpoint that serves one internal workflow a few hundred times a day is enormously more expensive per call than the same checkpoint on a shared deployment. Rent before you buy, and measure tokens per second on your own prompts rather than on a proxy workload.

Trying it before you bet a product on it

The honest first move on a model whose documentation is a day old is an evaluation, not a migration, and that is where a gateway does real work rather than rhetorical work. One key across 200-plus models at provider list price with 0% markup means the comparison you actually want — Nemotron 3.5 Super VL against the text or vision model you run today — is a change to a model string rather than a second vendor relationship, and automatic failover across providers means a workload you are only testing can run alongside production traffic without either depending on the other. That matters more here than usual: the two checkpoints behave differently enough on agentic work that the useful experiment is not "is this model good" but "which of these two builds, on which benchmark of mine, at what GPU cost."

The sequence that avoids wasting a GPU-month: pick the document-shaped workload first, because that is where the BF16-versus-NVFP4 gap is smallest and where the model's own numbers are best. Run it against the smaller checkpoint. Score it against an eval you wrote. Only then decide whether the agentic workloads — the ones where the quantised build loses 5.9 points on Terminal-Bench and 11.4 on WebVoyager — justify the 249 GB build and four GPUs.

What is still unanswered

Three gaps are worth watching because they change the decision rather than the marketing.

The first is independent evaluation. Every number in this article that comes from NVIDIA is labelled as such, and on this release that is all of them. No third party has published a throughput measurement, an accuracy reproduction or an evals harness comparison for either checkpoint. The 5x throughput claim and the paired quantisation table are internally consistent and clearly produced with care, and they remain vendor-reported and unreproduced. Treat them as an upper bound until somebody outside NVIDIA publishes the same experiments.

The second is the video path. The model accepts MP4 and the announcement leads with video understanding, but on the one video benchmark in NVIDIA's own table it lands at 32.85 while Qwen3.8-Flash-Next sits at 45.1. Training data covered video from 10,000 to a million hours, which is a wide range and tells you nothing about density or domain. If video is your use case, this is the row to test first and the row most likely to disappoint a procurement deck built from the announcement.

The third is the base checkpoint. Its publication under the same licence is the part of this release with the longest tail, because it lets a team with its own data run post-training rather than prompting around the vendor's. NVIDIA's own post-training corpus for this family is open too, and the effect sizes in NVIDIA's report run into the tens of points inside NVIDIA's harness. Nobody has yet published a credible independent fine-tune off that base. When someone does, it will tell you more about this model's real ceiling than the launch post does.

The weights are downloadable today and the licence lets you use them commercially. The two things you cannot download are an independent benchmark and a throughput figure with a named baseline — and both of those are what you would need before committing a production path to a model that is currently four days old.