Nemotron 3.5 Super vs GLM-5.5: one of these models exists

There is a real comparison between NVIDIA Nemotron 3.5 Super and GLM-5.5, and it is not the one the search results promise. Nemotron 3.5 Super — 120 billion total parameters with 12 billion active per token, a hybrid Mamba-2 and Transformer mixture of experts with a C-RADIO vision encoder, published under the fuller name NVIDIA Nemotron 3.5 Super VL and shipped on Ollama as code>nemotron-3.5-super/code> — is a set of weights that went up on Hugging Face on 12 October 2026, were announced by NVIDIA on 15 October, and can be downloaded and run tonight under the OpenMDW-1.1 licence. GLM-5.5 is a model that Z.ai has not released. There is no repository, no model card, no endpoint, no price and no parameter count. There are reports going back to June 2026 about a trillion-parameter launch, a leak write-up in October claiming it beats two frontier models, and silence from the lab.

Which means a head-to-head between these two would be one real model's specification against another model's reputation. That is not a comparison; it is a horoscope with benchmarks. So this page does the thing that is actually useful: it states what Z.ai has shipped, what is known about what it has not shipped, and then runs the comparison that a reader with a GPU and a workload can act on today.

What Z.ai has actually shipped

Z.ai's published flagship line runs through GLM-5, GLM-5.1, GLM-5.2, GLM-5.3 and GLM-5.3-Flash. On our own catalogue the highest GLM identifier we route is code>z-ai/glm-5.3/code>, alongside code>z-ai/glm-5.3-flash/code> and a free variant of the flash tier. There is no GLM-5.5 id, and the model list stops where the lab's log stops.

What GLM-5.3 is, from the figures we can verify: a 753-billion-parameter mixture of experts with 40 billion active per token, a 1,000,000-token context window with a 128,000-token output cap, text input, a licence Z.ai publishes under its own name with commercial use permitted, and a passed-through meter rate of $1.26 per million input tokens and $3.96 per million output. The flash tier sits below it at $0.15 and $0.50 with a 1M context and text, image and video input.

That flash tier is the more interesting competitor to this NVIDIA release than the flagship in one specific way, and there is a sibling page under this same batch that works through it. For the purposes of this article the point is narrower: if you want to compare Nemotron 3.5 Super to something Z.ai has shipped, you are comparing it to GLM-5.3, and that comparison is available today.

Why the GLM-5.5 reports are not usable numbers

The reporting around GLM-5.5 falls into three categories, and none of them will carry a purchasing decision.

The first is scheduling. Coverage from mid-2026 described a planned August launch aimed at ChatGPT and Claude, sourced to statements from Zhipu personnel and to industry chatter. The date passed without a release. A launch window that has already been missed is not evidence about a model.

The second is parameter count. Reports have circulated a trillion-parameter figure. Nobody has published a safetensors index, a config, or a per-layer expert count, which is what a parameter claim is normally checked against. The number may well turn out to be right; it is currently an assertion with no artefact behind it.

The third is benchmark claims. An October write-up claimed GLM-5.5 beats two named frontier models. There is no harness, no prompt set, no scoring configuration and no reproducibility path in that claim — which is exactly the gap that NVIDIA, whatever else one says about vendor-run evaluations, closed by publishing a NeMo Gym recipe with benchmark-specific containers and inference parameters alongside its table. A vendor benchmark with a recipe is weak evidence. A leak benchmark without one is not evidence.

There is a version of this article that treats the leak numbers as real, puts them in a two-column scoreboard and lets the reader assume the matchup is settled. That article would be worth nothing to anyone who has to make a decision, and it would be stale within a week of any actual announcement.

The comparison that can be made, run properly

Nemotron 3.5 Super VL against GLM-5.3, the flagship Z.ai has actually published.

• Total and active parameters — 120B total with 12B active, 512 routed experts per MoE layer · GLM-5.3 at 753B total with 40B active. On active parameters the NVIDIA model is a third of the size and should be dramatically cheaper per token to run on equivalent hardware, which is the entire design intent.

• Context — 1,000,000 tokens on both, with an important qualifier. NVIDIA's card says available memory and runtime settings determine the usable figure, and that prompts, images, history, tool results and generated tokens share the budget. Z.ai publishes a 128,000-token output cap for GLM-5.3 and NVIDIA publishes no output cap at all, which makes GLM-5.3 the more predictable model for long generations and the NVIDIA checkpoint the more open-ended one.

• Modality — text, image and video in with text out for the NVIDIA model, through a vision encoder and connector, with a 1.6 GiB projector on every published tag · text only for GLM-5.3, with the multimodal path living in the cheaper flash tier instead. If your input is a document page, the flagship GLM is not in the conversation.

• Metered price — none for the NVIDIA checkpoint, which is self-host only · $1.26 in and $3.96 out per million tokens for GLM-5.3, passed through at the provider's list price with no markup added.

• Third-party index, which is the only audited row in this list — the panel that scores these models places the NVIDIA checkpoint at roughly 26 and GLM-5.3 at roughly 45. That is an eighteen-point gap in favour of the Chinese flagship, measured by a third party rather than by either vendor, and it is the single most important number here.

• Release status — weights public on 12 October 2026, announced 15 October, four days of operational history · GA, with a shipped line of revisions behind it.

The gap on the index deserves to be stated without softening. NVIDIA's own published table cannot be used to argue it away, because the vendor did not run GLM-5.3 as a baseline — the GLM column on that table is GLM-5.3-Flash, the cheaper tier, and even there NVIDIA's model trails on most agentic rows. A 12B-active open checkpoint is not competing with a 40B-active flagship on general capability, and any page that tells you otherwise is selling something.

What the architecture choice actually buys you

The interesting question is not whether 12B active beats 40B active. It does not. The interesting question is what happens when you normalise for the hardware you can afford.

At its default 4-bit tag, the NVIDIA checkpoint is 82.3 GiB of weights plus a 1.6 GiB vision projector — one 96 GB card, with KV cache headroom that shrinks as you raise the context. GLM-5.3 at 753 billion total parameters does not fit on any single machine at any quantisation that preserves its behaviour, and its cheapest correct deployment is somebody else's cluster over an API at $1.26 and $3.96.

So the two models are not substitutes. One of them is a metered capability you rent and the other is a fixed cost you own, and the only thing that decides between them is how continuously you use the thing. Below a certain throughput, a $1.26-per-million-token flagship is simply cheaper than a 96 GB card and the electricity behind it. Above a certain throughput the card wins, and it wins without needing to be the better model.

The other thing the NVIDIA design buys, and it is not nothing, is the vision path on a permissive licence. GLM-5.3 is text-only. GLM-5.3-Flash takes images and video at fifteen cents and fifty cents per million tokens, which is remarkably cheap and is why the flash tier — not the flagship — is the right Z.ai comparison for a document pipeline. A self-hosted vision model at a fixed cost against a fifteen-cent hosted vision model is a genuine question with a genuine crossover point, and it is worth working out with your own page counts rather than with a leaderboard.

Three things about the NVIDIA checkpoint that are unresolved

Being even-handed in the other direction, because the openness of the NVIDIA release does not make its specification complete.

The default tag most people will pull is a 4-bit quantisation, and every published evaluation figure — NVIDIA's own table, the third-party panel score — is for BF16 or NVFP4. The version in actual use has no published evaluation. This is the largest unmeasured quantity on either side of this page.

The runtime's shipped sampling defaults are the card's text settings, temperature 1.0 with top_p 0.95, while the card's vision setting is temperature 0.6. A vision model whose default configuration is tuned for its text half is going to produce disappointing first impressions on scanned pages, and the fix is a config change rather than a different model.

And the specification disagrees with itself in two small places that matter to tooling: the Ollama config blob declares code>model_type: 123.6B/code> while every human-facing label says 120B, and the library page advertises a 1M context for its GGUF tags but no context figure at all for its three MLX tags. Neither is a defect in the weights. Both are things a deployment script will trip over.

Where this leaves a reader

If you are here because you wanted to know how Nemotron 3.5 Super compares to GLM-5.5, the answer is that the second model does not exist yet and any page that gives you numbers for it is inventing them. What exists is a leak with a parameter count, a missed launch window, and a claim with no harness behind it.

If you are choosing between what is actually shipped, the choice is between a metered Z.ai flagship that is meaningfully stronger on general capability and text-only, a metered Z.ai flash tier that is remarkably cheap and takes images, and an NVIDIA checkpoint you can own, that handles text, images and video, and that trails on the one independently measured score any of them has. Nemotron 3.5 Super is not on OrcaRouter — we route a little over two hundred models and there is no NVIDIA Nemotron entry among them — so if you want the NVIDIA checkpoint, the vendor's own channels and the local pull are your route.

GLM-5.3 and GLM-5.3-Flash are on our key, at the provider's list price passed through with no per-token markup, and a Z.ai price change lands on our side the same day rather than at your next contract renewal. That is the cheap experiment: run the workload against the model you can reach today, find out what it costs per unit of real work, and only then decide whether the fixed cost of a card you own beats a meter you rent.

And when GLM-5.5 does arrive — if it does — the version of this page worth reading will be the one written after somebody other than the lab has run it.