A generated title card reading 'Qwen4-Max vs Nemotron 3.5 Super' with the subtitle 'One you rent by the token, one you buy by the gigabyte' and three figure cards reading '43 vs 14 on the index', '$3.27 vs $0.56 per task' and '248.94 GB BF16 open weights'. A footer reads 'Open weights per NVIDIA; scores per Artificial Analysis Intelligence Index v5.0.1; hosted rates per Artificial Analysis.' Flat vector styling on a white background with blue-and-cyan gradient accents, the OrcaRouter mark composited bottom right.
Guides & Insights

Qwen4-Max vs Nemotron 3.5 Super: One You Rent by the Token, One You Buy by the Gigabyte

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The two models in this comparison are not competing for the same line of a budget. Qwen4-Max has been metered per token since 16 November 2026 at $1.50 per million input and $4.50 per million output in Singapore, and it is closed — no weights, no download, no self-hosting, ever. NVIDIA Nemotron 3.5 Super VL has been downloadable since 12 October 2026 under OpenMDW-1.1 as a 248.94 GB BF16 checkpoint or a 75.06 GB NVFP4 one, with 120 billion parameters of which 12 billion are active, and the hosted list price the independent tracker measures for it is $0.20 input and $0.50 output per million. On the index, Qwen4-Max scores 43 against Nemotron 3.5 Super VL's 14. On cost per completed task the tracker says $3.27 against $0.56. Read together, those four numbers describe a 29-point capability gap, a 92% price gap in the open model's favour, and a decision that is really about whether you want an endpoint or a file.

The licence is the first row, not the last

Most comparisons between a Chinese closed flagship and an NVIDIA open release treat licensing as a footnote. Here it determines what the other rows can even mean.

• Weights — Qwen4-Max: none published, and Alibaba has not said it will publish any for this tier, only that smaller Qwen4 models will be open-sourced on the same lineage. Nemotron 3.5 Super VL: BF16 and NVFP4 builds on Hugging Face, first created 12 October 2026, with NVIDIA's own announcement following on 15 October.

• Licence — OpenMDW-1.1 on the NVIDIA side, which permits commercial use and post-training on your own data. Alibaba's side is a commercial API agreement with per-token metering and nothing to accept beyond terms of service.

• Fine-tuning — possible on Nemotron 3.5 Super VL, and NVIDIA's card names customisation and domain adaptation as its intended use. Not offered at any price on Qwen4-Max; there is no fine-tuning endpoint in its capability listing.

• Deployment shape — one is a hosted endpoint in six regions that you reach through four wire formats. The other is a set of weights you have to find GPUs for, plus an inference stack, plus the operational work that goes with both.

• What you cannot do with either — buy Qwen4-Max's weights, or call Nemotron 3.5 Super VL with the assumption that NVIDIA will still be serving it next year. The second is the open-weights advantage stated negatively: nobody can retire weights you already have.

That asymmetry is why a straight index comparison is close to useless here. The 43-versus-14 gap prices capability, and the missing row prices optionality, and optionality is the thing the open model is actually selling.

Reading 14 against 26: the index moved under this model

There is a wrinkle worth naming, because it will confuse anyone comparing today's numbers to October's coverage. Artificial Analysis scored Nemotron 3.5 Super VL at 26 when it first evaluated it, and the current entry on Intelligence Index v5.0.1 reads 14. Both numbers are on the tracker's own pages; they are not comparable to each other, because the index revision changed underneath the model. What a reader should take from that is narrow: within v5.0.1, Nemotron 3.5 Super VL at 14 and Qwen4-Max at 43 are measured by one harness on one revision and can be placed on one line. The 26 is a historical artefact of a different rubric, and quoting the two side by side as if the model regressed by twelve points would be wrong.

The rest of the tracker's Nemotron entry is unusually informative for an open model, because hosting prices exist for it. It records $0.20 and $0.50 per million tokens, a 95% cache discount, output speed of 274 tokens per second, 13.44 seconds to first answer token, 5.0 minutes per index task, 1M tokens of context, and an average of $0.56 per completed index task on 119 million generated output tokens. The description is candid in both directions: among the leading models in intelligence for open-weight models of similar size, slower than average per task, very verbose.

Qwen4-Max's column has no speed, no latency and no time per task — the performance card reads N/A. It has 43, a rank of 38 of 81, $1.50 and $4.50, and 127 million output tokens. The verbosity is comparable to Nemotron's, which makes the cost-per-task difference almost entirely a price difference rather than a behaviour one: at roughly a ninth of the per-token rate on output, the open model finishes tasks at a sixth of the cost while scoring 29 points lower.

A generated two-column scoreboard titled 'Qwen4-Max vs Nemotron 3.5 Super VL — the scoreboard'. The Qwen4-Max column reads Distribution hosted API only, Weights none published, AA Index 43 rank 38 of 81, Price per index task $3.27, Parameters 2.4T total 71B active, Fine-tuning not offered. The Nemotron 3.5 Super VL column reads Distribution open weights OpenMDW-1.1, Weights 248.94 GB BF16 / 75.06 GB NVFP4, AA Index 14, Price per index task $0.56, Parameters 120B total 12B active, Fine-tuning permitted. A footer reads 'Open weights per NVIDIA; AA figures per Artificial Analysis v5.0.1; hosted rates per AA.'

What 120B with 12B active actually buys you

Nemotron 3.5 Super VL is a mixture-of-experts with 120 billion total parameters and 12 billion active per token, which is the number that governs inference cost when you host it yourself. It takes text and images and returns text. Its language backbone was initialised from Nemotron 3 Super Base, a text-only model, with the vision path added afterwards — a C-RADIO encoder and a connector installed as a distinct training stage after foundation training, distillation, Indic continued pre-training, long-context extension and software-engineering mid-training had all completed. Qwen4-Max is the opposite construction: a 2.4-trillion-parameter backbone with 71 billion active per token, native vision from the start, 328 billion further N-gram embedding parameters looked up rather than multiplied, and a context path built on three Gated DeltaNet layers compressing history for each Qwen Sparse Attention layer.

One of those architectures was designed around images. The other had images grafted on. Whether that difference shows up in your workload is an empirical question, but the direction of the risk is knowable: a retrofitted vision path is where a document-extraction pipeline finds its first regression, and the vendor's own card gives a hint in the same direction, with an IFBench score of 70.86 on loose prompts against 82.9 for the larger Nemotron 3 Ultra.

The hosting arithmetic is the other half. NVIDIA publishes two checkpoints: 248.94 GB of BF16 and 75.06 GB of NVFP4, a 173.88 GB gap for roughly a third of the storage and memory footprint. The NVFP4 build is the one that fits on a single node of modern accelerators rather than a cluster, and it is NVIDIA's own checkpoint rather than a third-party quantisation, which matters if you care about reproducing the vendor's published numbers. Using 12 billion active parameters per token, the throughput question becomes a matter of your batch size and your GPU, not of a rate card — which is exactly the freedom the open licence buys and exactly the work it costs.

A headless-browser screenshot of the Hugging Face model card for nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16, captured 16 November 2026 in English, showing the repository name and like count, Image-Text-to-Text, Transformers, Safetensors and PyTorch tags, the nemotron_h_omni and nemotron-3.5 tags, a link row for the pre-training datasets, developer page and paper, and the start of the card describing it as the full-precision release of NVIDIA Nemotron 3.5 Super VL, a vision-language model for reasoning.

When the 29-point gap stops being the deciding number

Three workloads make the open model the better purchase despite scoring less than a third of Qwen4-Max.

• Regulated or air-gapped deployments — where the question is not price but whether data can leave the building, and a rented endpoint is disqualified before the index is consulted. Nemotron 3.5 Super VL at 75.06 GB NVFP4 is a plausible on-premise model; Qwen4-Max has no on-premise form.

• High-volume classification, routing and extraction — where 14 on a general index can still be sufficient, output lengths are pinned by a schema, and 274 tokens per second at $0.20 per million input is an order of magnitude cheaper than anything metered at frontier rates.

• Anything that needs to learn your data — a fine-tune on your own images and documents is available on one side and unavailable on the other at any budget, and no amount of prompt engineering substitutes for it when the domain vocabulary is yours alone.

And three that make Qwen4-Max the right answer anyway: long-horizon agentic work where the score gap translates directly into failed multi-step tasks, requests that cross a few hundred thousand tokens and need a single-request context with no chunking, and any team that wants structured output, parallel tool calling and caching from an endpoint today rather than an inference stack next quarter. Qwen4-Max also publishes both of its input ceilings — 991,808 tokens in non-thinking mode, 983,616 in thinking mode — which is the kind of number a self-hosting team spends a week discovering.

What neither of these two is, and what is actually routable

Neither model is in the OrcaRouter catalogue as of 16 November 2026. Qwen4-Max is reachable through Alibaba's own Model Studio API and several third-party platforms; Nemotron 3.5 Super VL is downloadable weights plus a handful of hosted providers. A comparison page that implied either was one API key away would be lying about the effort involved.

What is routed here is the prior generation of the closed side: qwen/qwen3.8-max at $2.00 in and $6.00 out per million tokens, the vendor's list price passed through with 0% markup, on the same key and the same OpenAI-compatible endpoint as the rest of the catalogue. For a team evaluating this matchup, that is the useful intermediate position — measure your workload against a routed Qwen flagship you can call today, and use that number to decide whether the Qwen4-Max step up is worth a second contract, rather than buying the integration before you have the evidence.

A headless-browser screenshot of the OrcaRouter model page for qwen/qwen3.8-max, captured 16 November 2026 in English, showing the Qwen3.8 Max heading and identifier, text, image and video input with a 1M-token context window, an OpenAI-compatible endpoint at api.orcarouter.ai/v1, the price block reading $2.00 input and $6.00 output per million tokens, measured latency, 144.9M tokens served and a zero-data-retention note.

The path that usually decides it in practice is not the flagship at all. A high-volume pipeline tends to end up split: a cheap model doing the sorting and extraction, a strong model doing the final pass, and a fallback for the day the primary has a bad afternoon. OrcaRouter's routing DSL exists to express that split in one call rather than three integrations, and automatic failover exists to make the cheap half safe to depend on. Neither of the two models in this piece is on that path yet; the pattern is worth knowing because it is where a matchup like this one usually lands.

The test that settles it, and what it costs

Run 200 of your own hardest tasks through both. Not the vendor benchmarks, not the index — your tasks, with your output lengths.

Price the Qwen4-Max arm at $1.50 and $4.50 per million on measured tokens, and note that the model has no published speed, so measure latency yourself or accept that you are buying an unknown. Price the Nemotron arm two ways: at the hosted $0.20 and $0.50 if you want the cheap comparison, and at your own GPU-hour cost divided by tokens served if you are serious about self-hosting, because the hosted rate is somebody else's utilisation curve and will not be yours. Then look at the failures rather than the averages, since a 29-point index gap shows up as tasks that do not complete at all, not as tasks that complete slightly worse.

The answer that most teams arrive at is not a winner but a split: the open model on the volume and the regulated data, the metered flagship on the work where a partial answer is worthless. That is a more expensive architecture than either single choice, and it is usually the correct one — which is also why the licensing row deserves to be read first and the index row second, and why the $0.20 hosted rate for a model nobody has to retire is the most interesting number in this comparison.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily