Nemotron 3.5 Super VL vs Kimi K3: A Metered Flagship Against a Weight Release With No Rate Card

The shortest honest answer is that only one of these two models has a price, and it is not the one from NVIDIA. Kimi K3 is sold as a metered service at $3 per million input tokens and $15 per million output, with a 90% cache discount, and it is the model you can call today from a single key. NVIDIA Nemotron 3.5 Super VL is a 120-billion-parameter, 12-billion-active open-weight checkpoint under OpenMDW-1.1 with no published rate card at all — you either run it on your own accelerators or you negotiate with whoever is serving it. Its weights went up on Hugging Face on 12 October 2026 and NVIDIA's own announcement followed on 15 October, which makes this a three-day-old release rather than a mature product. That asymmetry, not the benchmark table, is what decides the matchup.

What each of these actually is

NVIDIA Nemotron 3.5 Super VL is a hybrid vision-language model: interleaved Mamba-2 layers, a mixture-of-experts with 512 routed experts per MoE layer, and select attention layers, paired with a C-RADIO vision encoder and a vision-language connector. It carries 120B total parameters and activates 12B per token, supports up to 1M tokens of context, and takes text, images and video in with text out. Native Multi-Token Prediction gives it speculative decoding without a separate drafter checkpoint. NVIDIA ships a BF16 reference build and an NVFP4 build, with deployment recipes for 4×H100 or 2×GB200 at BF16. Pre-training and continued pre-training drew on a corpus NVIDIA describes in the trillions of tokens, and post-training added synthetic data plus reinforcement learning across dozens of published environments.

Kimi K3 is Moonshot AI's flagship: a 2.8-trillion-parameter mixture-of-experts with 104B active parameters, a 1M-token context window, and native image input alongside text. It is served rather than merely published, it exposes a top-level reasoning_effort control instead of sampling parameters, it speaks the OpenAI API shape, and it is positioned for long-horizon coding and knowledge work inside agent harnesses.

• Scale — Super VL activates 12B parameters per token from a 120B pool. Kimi K3 activates 104B from a 2.8-trillion-parameter pool. That is roughly a nine-fold difference in per-token compute.

• Context — 1M tokens on both, and both are designed to hold a working session rather than a single prompt.

• Modalities — Super VL takes text, still images and sampled video frames; Kimi K3 takes text and images. Video is the one input axis where the two are not equivalent.

• Price — no published rate anywhere for Super VL as of 16 October 2026; Kimi K3 is $3 in / $15 out per million tokens with a 90% cache discount.

• Licence and control — OpenMDW-1.1 with commercial use, open weights, open datasets and training recipes for Super VL; a commercial API with no weight download for Kimi K3.

• Deployment floor — 4×H100 or 2×GB200 at BF16 for Super VL, or the vendor's NIM microservice; nothing to deploy for Kimi K3.

The eight-point gap that decides the generalist case

On the Artificial Analysis Intelligence Index, Kimi K3 sits at 44 and NVIDIA Nemotron 3.5 Super VL at 26 — both non-estimated, both measurable by a third party rather than the labs that built them. Eighteen points is not a rounding error; it is roughly the distance between a frontier generalist and a mid-tier one. Kimi K3 is also slower and more expensive per token than the models around it, which the same panel notes plainly.

That number is the whole argument for anyone whose workload is text-only reasoning delivered over an API, and it is worth saying so without hedging. If your agent's job is to read a repository, plan a change and hold a long conversation about it, K3 is the stronger model and it is available now at a published price.

Where the comparison stops being about intelligence

NVIDIA did not build Nemotron 3.5 Super VL to win a general reasoning index, and reading the matchup as if it had is how people end up disappointed. The release is aimed at what the announcement calls always-on agents: high-volume, long-running, document-and-screen-shaped work that a company wants to run inside its own boundary, post-train on its own data, and audit. The vendor's own framing is efficiency — the claim is that Super VL is five times faster at task completion than the next best model in its class, and that it reaches accuracy comparable to NVIDIA's largest reasoning model at roughly one-fifth the parameter count.

Those are vendor claims, measured on NVIDIA's own harness, and nobody outside NVIDIA has reproduced them yet. The honest reading after one day of public availability is: the architecture is documented, the weights are downloadable, the licence is permissive, and the throughput figure is an assertion until an independent party reruns it.

What is checkable is the flavour of work. Super VL's published strengths are optical character recognition, tables and charts, structured information in visual documents, and multistep business workflows — insurance claims, credit documents, engineering drawings, search across application silos. That is a different job from writing a multi-file patch, and it is a job where a 12B-active model that finishes a page-fast task is worth more than a 104B-active model that finishes it better and later.

What one key does with each of them

This is where the asymmetry bites hardest, because the two models have opposite access stories.

Kimi K3 is on our catalogue as kimi/kimi-k3 at Moonshot's own list price, with 0% markup — the provider's rate passed through untouched, so when Moonshot changes the number, the number you pay changes the same day. It arrives through the same OpenAI-compatible endpoint as everything else, which means swapping it in is a model-string change and not an integration project. Because Kimi K3 exposes reasoning_effort rather than temperature and top_p, that same endpoint lets you dial reasoning depth per request without a second code path.

NVIDIA Nemotron 3.5 Super VL, by contrast, is not something we serve. It is a checkpoint you host, and the cost model is capital and utilisation rather than tokens. What we can do for a self-hosted model is the surrounding plumbing: automatic failover across upstream providers so that a spike or an outage on a hosted model degrades gracefully instead of failing, and the routing DSL that composes several models into one call — send the mechanical extraction to a small cheap model and escalate only the pages that need judgement. That pattern is exactly what Super VL's economics invite, because a 12B-active model is cheap enough to run on everything and a 104B-active one is not.

The realistic production shape for the next few months is therefore not a choice. It is K3 for the reasoning-heavy branch and a self-hosted Super VL for the document-volume branch, with one router in front deciding which branch a request belongs to.

What would change the answer

Three things. First, an independent reproduction of the five-times claim — if Super VL really completes its class of tasks at a fifth of the wall-clock, the case for running it inside a regulated perimeter gets much stronger, and if it does not, the accuracy-per-token argument collapses back to a normal open-weight one. Second, a published rate from a serving provider: the moment Super VL appears under a metered API, the deployment-floor objection disappears for everyone who does not own eight H100s, and it becomes a straight price-versus-index question where K3 currently wins on index and loses on nothing. Third, Kimi K3 itself is not standing still — Moonshot has been iterating fast, and a 3-series release would reset this page rather than refine it.

Until one of those lands, the split is clean: Kimi K3 if you want the most capable model per token today and can pay for it; NVIDIA Nemotron 3.5 Super VL if you want a document-and-screen agent you own, post-train and keep on premises, and are willing to be the first to measure what it does.