Nemotron 3.5 Super VL: NVIDIA's 120B Open Vision Model Ships, and the 5x Claim Has No Scoreboard Yet

NVIDIA Nemotron 3.5 Super VL shipped on October 15, 2026 — an open 120-billion-parameter vision-language model with open weights, open datasets and training recipes, and a claim attached to it that nobody can check yet. The company's launch post calls it "the fastest open agentic model in its class — 5x faster than the next best," and attributes the number to the Artificial Analysis Intelligence Index leaderboard. As of this writing, Artificial Analysis does not list Super VL at all; its Nemotron 3.5 entries stop at Nemotron 3.5 Lightning. That does not make the claim false. It makes it vendor-reported, and it should be read that way until a third party runs the checkpoint.

The other thing worth knowing up front is a three-day gap. The weights went up on Hugging Face at 2026-10-12T15:32Z — the BF16 post-trained build, the base model and the NVFP4 build all within a minute of each other — and the announcement post did not appear until October 15 at 15:00Z, with the model card's own release-date field still reading 10/15/2026. NVIDIA uploaded first and announced later. That is now the normal shape of an open-weights release: the repository is the primary source, and the blog post is the commentary.

What actually shipped, and what did not

Three Super VL checkpoints, all live now:

• NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 — the full-precision post-trained model, the reference build and the one to fine-tune from.

• NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 — the quantized deployment build, produced with NVIDIA Model Optimizer and quantization-aware distillation from the BF16 model.

• NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16 — the pre-post-training base, for anyone who wants to run their own post-training rather than adapt NVIDIA's.

There is a fourth repository, NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr, which is not a model you prompt. It is a speculative-decoding draft checkpoint: load it alongside the NVFP4 target and it proposes tokens ahead of the main model. Its own card dates it 10/15/2026 and describes it as intended for LiLiCorr deployments tuned for DGX Spark and low-concurrency data-center work.

Two things did not ship, and each matters. There is no price. NVIDIA's announcement states the model is available as an NVIDIA NIM microservice — a production microservice the company says is refreshed every two weeks with security patches and performance work — but publishes no per-token figure anywhere, for NIM or for the third-party platforms carrying it. And there is no independently measured speed number, for the reason above. Everything in the next two sections is NVIDIA's own reporting.

The architecture in one paragraph, because it explains the deployment story

Super VL is a Mamba-2–Transformer hybrid mixture of experts: interleaved Mamba-2, MoE and attention layers, 512 routed experts per MoE layer, plus a C-RADIO vision encoder and a vision-language connector bolted onto the language backbone. 120B parameters total, 12B active per token. Context runs up to 1M tokens, and Multi-Token Prediction is built in natively, which is what the LiLiCorr draft checkpoint exploits. Inputs are text, images and video — video as MP4 up to two minutes, sampled as frames — and documents can be fed as page images with accompanying text. Output is text only.

That shape is why the deployment table is the most useful part of the documentation. NVIDIA's BF16 build wants 4× H100 or 2× GB200. The NVFP4 build wants 1× DGX Spark, 1× H200, 2× H100, or 1× GB200. The gap between those two lines is the whole reason the quantized checkpoint exists, and it is the difference between "a 120B model you rent" and "a 120B model on the desk."

The benchmarks NVIDIA published, and the two problems with them

Label all of this vendor-reported. NVIDIA evaluated Super VL through NeMo Gym and published the recipes, containers, prompts and scoring settings — which is more than most labs do — but NVIDIA trained the model, chose the baselines and ran the harnesses. Nothing here has been reproduced outside the company.

The agentic numbers are the ones the launch leans on. SWE-bench Verified 71.10, SWE-bench Multilingual 67.72, SWE-bench Pro 51.78, tau2-bench retail 69.24, airline 76.50, telecom 73.19, BrowseComp 78.91, and GraphWalks 89.71 at the over-1M-token setting. On reasoning, GPQA-Diamond 89.52 and HLE without tools 31.79, rising to 46.01 with tools — that last figure, NVIDIA notes, comes from a separate internal harness rather than the standard path.

On vision, which is the new capability relative to the text-only Nemotron 3.5 Lightning from August: CharXiv-RQ 80.82, MathVision 78.16, V* 85.60, OCR-Reasoning 65.08, OCRBench-v2 at 68.12 English and 59.15 Chinese, MMLongBench-Doc 62.82, OSWorld 52.14, and Video-MME-v2 32.85.

Now the two problems.

First, the Terminal-Bench 2.1 figure is stated twice and does not agree with itself. The model card reports 66.70. The technical report's summary table reports 60.67. Everything else in those two documents lines up, so this is not two different models — it is one number carrying two values in NVIDIA's own paperwork, and until the recipes are re-run it is not safe to quote either. That is the sort of thing a single independent replication settles in an afternoon, which is exactly why the absence of one matters this week.

Second, the headline efficiency claim and the headline intelligence claim pull in opposite directions, and only one of them is checkable. NVIDIA says Super VL reaches accuracy on multimodal agentic tasks "comparable with NVIDIA's largest reasoning model at roughly one-fifth the parameter count" — Nemotron 3 Ultra 550B A55B has 55B active parameters against Super VL's 12B, a ratio of 4.6×, and the report says Super VL's 51.78 on SWE-bench Pro does surpass Ultra. Fine.

But look at the columns where Super VL is genuinely behind. MMLU-ProX 76.91, against Qwen3.5-122B-A10B at 83.2 and Qwen3.8-Flash-Next at 83.0. AA-Omniscience accuracy 26.20, which is close to the floor of that benchmark and sits below DeepSeek-V4-Flash-0731's 45.0. Video-MME-v2 32.85 against Qwen3.8-Flash-Next at 45.1. BabyVision 29.60 against 46.8. Agentic-terminal at 60.67 or 66.70 against 85.80. Super VL is not the accuracy leader in this weight class on most rows; it is a 12B-active model that NVIDIA is selling on speed and on the breadth of what it ingests in one call, not on beating everything at everything.

The efficiency numbers, which are the actually interesting ones

Strip out the intelligence comparisons and the engineering numbers are concrete:

• Speculative decoding — NVIDIA's post-training MTP boosting lifts average acceptance length from 4.287 to 4.718 tokens under greedy decoding and 3.787 to 4.612 under probabilistic draft sampling.

• Quantization — the released NVFP4 checkpoint recovers 99.62% of BF16 accuracy as a median across 36 benchmarks, so the deployment build is close to lossless by NVIDIA's own measure.

• Throughput — up to 4,856 output tokens/s/GPU on a single GB200 on a 20-turn agentic proxy workload while sustaining at least 50 tokens/s/user; NVFP4 at 2.43× BF16 throughput under the same conditions.

• Relative throughput — 2.34× Qwen3.5-122B-A10B and 1.90× Qwen3.8-Flash-Next on that workload.

• Serving stack — on NVIDIA Dynamo, NVIDIA claims up to 8× more concurrent users at 85% lower latency.

Every one of those is a vendor number. The one that is easiest for anyone to falsify is the 4,856 tokens/s/GPU figure, because it is a property of a specific GPU and a specific workload, not of a benchmark run behind a scoring harness.

Serving it, and what the two builds cost you in hardware

Both vLLM and SGLang are named as supported runtimes, and NVIDIA ships working launch configurations: four-way tensor parallelism with three MTP draft tokens on 4× H100 for BF16, two-way with five draft tokens on 2× GB200. Hardware coverage is Hopper and Blackwell, Linux only, and NVIDIA lists DGX Spark, DGX Station, RTX PRO, RTX and Jetson on the client side — the same silicon spread the model is positioned against.

The practical decision for a team reading this on launch day is which checkpoint to pull. If you are fine-tuning, the BF16 build is the only sensible starting point, because the NVFP4 build has already been through quantization-aware distillation and further training will not compose cleanly with it. If you are serving, the NVFP4 build is the one that fits on a single H200 or a DGX Spark, and a 99.62% median recovery makes it the default unless you have measured a specific regression. Pulling the base model only makes sense if you intend to write your own post-training pipeline, which is a different project.

What Super VL is actually for

Read NVIDIA's own framing and the target is narrow and specific: always-on enterprise agents that digest visual documents. The announcement's example is a web page — extract the text, read the charts, return findings — and the same claim is extended to PDFs, tables and structured documents. The argument for one model instead of a text model plus a separate vision pipeline is context continuity: the document, the data and the decision stay in the same window rather than being handed between systems.

That argument is credible on the spec sheet because the context window is 1M tokens and the vision encoder is trained jointly, not bolted on after the fact. Whether it survives contact with real documents at 1M tokens is exactly the question the GraphWalks and MMLongBench-Doc numbers are aimed at, and both are vendor-reported.

NVIDIA also lists a long adoption roster in the launch post — Scale AI building agents for Skydance and EY, AfterQuery for Hebbia's credit-agreement analysis, AI21 and Alice on cybersecurity, Mercor on coding agents, plus Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys among integrators. Those are partnerships named in a press release, which is a statement of intent rather than evidence of production traffic. Treat the roster as a signal about which verticals NVIDIA is targeting, not as a reliability claim.

How to try it without betting a production path on it

The honest position this week is that Super VL is an interesting, well-documented, entirely unverified open model, and the right way to evaluate an unverified open model is not to route production traffic to it on day one.

Locally, the NVFP4 build on a single H200 or a DGX Spark is the cheap test: pull the checkpoint, run your own document set through it, and compare against whatever vision model you use today. That costs you GPU time and nothing else, and it is the only way to get a number that is not NVIDIA's. NVIDIA's own documentation gives you one free head start: the model card and the technical report both specify temperature 1.0 and top_p 0.95 for text, and temperature 0.6 with top_p 0.95 for vision — a starting point you would otherwise spend a day rediscovering.

If you would rather test it behind an API, the availability line matters: Super VL is on Hugging Face, as NVIDIA's own NIM microservice, and through several third-party platforms. NVIDIA has published no price for any of those paths, so any per-token figure you see this week is set by the platform, not by the vendor. On the OrcaRouter side, Nemotron 3.5 Super VL is not on our catalogue — we do not serve it today, so there is nothing here to try it through yet, and we would rather say that than imply otherwise. Everything we do serve sits behind one key across 200-plus models at provider list price with 0% markup, and where a model is available from more than one provider, automatic failover across providers gives you a way to route to a new model on a fraction of traffic and let it fall back to the incumbent when it stumbles. That is the mechanism this kind of launch wants; it is not an offer to serve this particular checkpoint.

One more piece of the release worth a look regardless of whether you want the model: NVIDIA published its post-training datasets alongside it — spans covering agentic tool use, finance, instruction following, long context, fairness and vision-language understanding — plus the RL environments in NeMo Gym. The dataset release is often more reusable than the checkpoint, and it is the part of this launch that costs NVIDIA the most to give away.

What to watch

Four things would change this article, in rough order of how much they would matter.

A first independent benchmark, from Artificial Analysis or anyone else, that either produces the Super VL row the 5x claim implies or declines to. Until that exists, the launch's central efficiency claim is a sentence in a press release with a leaderboard's name attached to it, and the caveat belongs in the same breath as the number.

A price. NVIDIA has told you the hardware, the runtimes and the licence, and withheld the one figure a buyer needs to compare this against a closed model. When NIM pricing appears, the comparison becomes arithmetic instead of positioning.

The Terminal-Bench 2.1 discrepancy — 66.70 on the model card against 60.67 in the technical report. Whichever is right, one of NVIDIA's two primary documents is wrong, and a corrected re-run would settle several other rows at the same time.

And the walk from "available now" to "actually served." Super VL's benchmark rows compare it against Qwen3.5-122B-A10B, Qwen3.8-Flash-Next, DeepSeek-V4-Flash-0731, GLM-5.3-Flash and Nemotron 3 Ultra 550B A55B — all of which are mature enough to have measured throughput, real prices and production track records. A 12B-active model that ingests video, 1M-token documents and tool calls in one pass is a genuinely useful shape. Whether it earns a routing rule is a question about latency and price stability under load, and NVIDIA has published the first of those only on its own hardware, under its own workload, in its own units.

Benchmarks age and prices move; the failure record does not: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.