Nemotron 3.5 Super VL vs Qwen3.8-Max: The Open-Weight Tier Against the Metered Frontier

Someone planning a document-processing pipeline this quarter has exactly two shapes to choose between, and they are not the same shape. NVIDIA Nemotron 3.5 Super VL is a 120B-parameter, 12B-active vision-language checkpoint under OpenMDW-1.1 whose weights went up on 12 October 2026, three days before NVIDIA's own announcement on 15 October: you download it, you host it, you pay in GPUs and electricity. Qwen3.8-Max is Alibaba's proprietary flagship — a 2.4-trillion-parameter mixture-of-experts with 95B active parameters, vision input, a 1M-token context window and per-token billing at $2 per million input and $6 per million output with an 88% cache discount. One is capital expenditure with full control; the other is an operating expense with a rate card. Everything else in this comparison is downstream of that split.

Two business models wearing a benchmark comparison

It is tempting to line these up on capability and declare a winner. That framing hides the actual decision, which is about who owns the failure mode. With Super VL, if throughput disappoints you, you retrain, re-quantise or buy different hardware — the ceiling is yours to raise. With Qwen3.8-Max, if throughput disappoints you, you change the model string or you negotiate; the ceiling belongs to Alibaba, and so does the roadmap.

That difference is legible in what each vendor published. NVIDIA shipped a technical report, training recipes, open datasets, RL environments and three separate speculative-decoding paths for the family — the apparatus of a model meant to be deployed and modified. Qwen shipped an API with a price list and a cache discount, which is the apparatus of a model meant to be consumed.

What Super VL gives you that a hosted endpoint cannot

NVIDIA Nemotron 3.5 Super VL is a hybrid: interleaved Mamba-2 layers, a mixture-of-experts with 512 routed experts per MoE layer, and select attention layers, joined to a C-RADIO vision encoder and a vision-language connector. Text, images and video frames go in; text comes out. Its language backbone was initialised from Nemotron 3 Super Base, distilled for a trillion tokens from Nemotron 3 Ultra Base, then given a 35-billion-token extension to 1M context, software-engineering mid-training and a long-context recovery stage. Native Multi-Token Prediction supplies speculative decoding without a separate drafter. It runs at BF16 on 4×H100 or 2×GB200, with an NVFP4 build and a NIM microservice for production deployment.

• What you own — the weights, the tokeniser, the vision encoder, the post-training pipeline and the audit trail. Every prompt stays inside your perimeter, which for insurance, credit, health or defence work is not a preference but a requirement.

• What you pay — accelerators. Four H100s at full precision is the entry ticket, and the cost curve is utilisation: expensive when idle, near-free at saturation.

• What you give up — a rate card. As of 16 October 2026 there is no published price for Super VL from anyone, so any total-cost comparison you build today rests on your own throughput measurements, not a vendor number.

What Qwen3.8-Max gives you that a checkpoint cannot

Qwen3.8-Max activates 95B parameters per token from a 2.4-trillion-parameter pool, takes images as well as text, and holds a context window just under a million tokens by default. On Artificial Analysis it carries an Intelligence Index of 45, which is nineteen points above the 26 recorded for NVIDIA Nemotron 3.5 Super VL on the same index — and unlike most of the numbers on this page, both of those are third-party measurements rather than vendor claims.

• What you get — a frontier-class generalist, vision included, callable this afternoon with an API key. No hardware, no ops, no capacity planning.

• What you pay — $2 in and $6 out per million tokens, with an 88% discount on cached input, which changes the economics of any workload with a large stable prefix: a long system prompt, a fixed document, a persistent tool schema.

• What you give up — the weights. There is no download, no local deployment, no fine-tune of the serving model, and no ability to keep data on your own metal.

The cache discount deserves a specific note because it is the most commonly mis-modelled number in either column. An 88% reduction on cached input does not mean an 88% cheaper bill; it means the repeated part of your prompt costs roughly a ninth of list, and the benefit scales with how much of each request is genuinely identical to the last. Pipelines that resend a large static preamble benefit enormously. Pipelines that send a fresh document every call benefit barely at all.

The price line, run honestly

The instinct is to compute a crossover — at how many tokens does self-hosting win? That calculation produces a number and the number is usually wrong, because it omits the two costs that dominate: engineering time to keep a self-hosted inference stack current, and the opportunity cost of a smaller model on the hard cases.

A more useful version: run your real traffic profile against Qwen3.8-Max's published rates and get an exact monthly figure, because that side has a rate card. Then ask what fraction of that traffic a 12B-active model can handle to your standard, and price only that fraction in GPUs. In most document-extraction pipelines the answer is that the bulk of the volume is mechanical — invoices, forms, charts, repeated layouts — and a small tail carries the judgement. The bulk is where self-hosting wins; the tail is where a frontier model like Qwen3.8-Max wins.

That split is what our own routing layer exists to serve, and the relevant facts are narrow enough to state plainly. Qwen3.8-Max is on our catalogue as qwen/qwen3.8-max/ at Alibaba's list price with 0% markup — whatever the vendor charges, that is what you pay, and a vendor price cut lands the same day rather than at the next billing cycle. NVIDIA Nemotron 3.5 Super VL is not on our catalogue and never will be in the form of an endpoint, because it is a checkpoint you host rather than a service anyone can resell you. What sits between the two is composition: the routing DSL lets a single call fan out, so mechanical pages can go to whatever cheap model you point at them while the ones that need judgement escalate, and automatic failover moves a request to a healthy upstream when the one it was heading for degrades. One key for the hosted half, your own GPUs for the other, one policy deciding which request goes where.

A worked shape rather than a hypothetical: the same pipeline can send a one-page form to a small text model through the catalogue, route a chart-heavy annual report to a vision model, and hold Qwen3.8-Max in reserve for the minority of documents where the extraction is genuinely ambiguous. The reserve model costs $6 per million output tokens and nothing when unused, which is a far better arrangement than provisioning a fourth H100 for a workload that spikes twice a month.

Who should do which, starting this week

If your work involves regulated data — any category where sending a page to a third party requires a legal review — the decision is made for you and it is Super VL, hardware floor and all. If your work is high-volume document and screen understanding under a normal privacy posture, the cheapest correct answer for most teams right now is the metered frontier model, because it works today and its price is known. If you are somewhere in between, which is where most real pipelines sit, the two are not competitors at all: Qwen3.8-Max for the hard tail and a self-hosted Super VL for the volume, with a router deciding per request.

The thing that would change this page is a published rate for Super VL from any serving provider. The moment a Nemotron checkpoint appears behind a metered endpoint, the deployment-floor objection evaporates, the comparison becomes price against price, and the interesting question shifts from which model to which provider. Until then, the two columns are still answering different questions — and the honest answer to "which is better" is that a model with a price and a model with a licence are not the same kind of object.