GLM 5.5: What the Name Is, and What Z.ai Actually Ships Today

GLM 5.5 is a version number, not a model. Z.ai (Zhipu AI) has never put it on a shipped release, and as of 13 October 2026 there is no model card, no API identifier, no pricing row, no weight repository and no licence file anywhere in the line carrying it. What does exist behind the name is the line itself: GLM-5.3, Z.ai's text flagship, and GLM-5.3-Flash and GLM-5.3-FlashX, the natively multimodal models that followed it in August 2026 — all three released, priced on Z.ai's own rate card, and downloadable from the lab's own Hugging Face organisation. So the useful answer to "what is GLM 5.5" has three parts, and only the third one is about a model: what the name refers to, what would have to appear before it became a model, and what a reader who wants a shipped model right now should actually run.

This page exists because the question is being typed and nothing answers it. On our own property, queries beginning "glm 5.5" drew 1,415 impressions and 64 clicks across nine queries in the twenty-eight days to 12 October 2026, and 1,312 of those impressions — 93% — were on the bare name with no qualifier at all. Readers are not exploring variants of this term. They are typing it once and hoping for a definition, and today they land on speculation or on a comparison whose opponent does not exist.

The eleven pages that already own the name

Before writing another word about it, worth being precise about what is already published, because the crowded stem is the reason this page is a reference rather than a news piece. Eleven slugs carry the GLM 5.5 name on this site today, all verified live:

• glm-5-5-leak — the August report on an anonymous model whose serving fingerprint matched Z.ai's stack, with the working guess that it might be GLM 5.3-Vision or GLM 5.5.
• glm-5-5-release-date — the standing tracker for whether and when the name becomes real.
• Nine comparisons — against claude-opus-5, claude-sonnet-5-5, gemini-4-argon, gpt-6, mistral-large-4-0, muse-wonder, reflection-beam, step-5-preview, deepseek-v4-pro, gemini-3-1-pro, grok-4-5, kimi-k3, minimax-m3 and qwen-3-8 (several of these pair the name with more than one opponent, and the older generation of them sits against models that have since been superseded).

That is a page count, not a coverage count, and the difference matters. Two of the eleven are speculation and nine are comparative, which means a reader who searches the bare name gets either a countdown or a head-to-head — and in every head-to-head one of the two sides has no specification, no price and no endpoint, so the comparison can only really be run against the shipped models. What none of the eleven does is answer the question the 1,312 bare impressions are asking. That is this page's job, and it is not going to become the twelfth speculative page to do it: no leak recap, no release countdown, no expectations.

Nine surfaces, checked on 13 October 2026

A claim that something does not exist decays the moment it is written, so here is the method rather than only the conclusion. Every line below was re-run today and every one is repeatable by a reader in an afternoon.

• Z.ai's release log at docs.z.ai/release-notes/new-released — the newest entry is GLM-5.3-Flash, dated 26 August 2026, above GLM-5.3 (18 August), GLM-5.2 (16 June), GLM-5.1 (7 April) and GLM-5 (12 February). Nothing numbered 5.4 or 5.5 sits above it.
• The documentation index at docs.z.ai/llms.txt — the model pages it enumerates are GLM-5.3-Flash/FlashX, GLM-5.3 and GLM-5.2. There is no GLM-5.5 page to fetch.
• docs.z.ai/guides/llm/glm-5.5.md returns 404, as does z.ai/blog/glm-5.5.
• Z.ai's pricing page — the rate card carries GLM-5.3-Flash at $0.15 / $0.50, GLM-5.3-FlashX at $0.37 / $1.25, GLM-5.3 and GLM-5.2 at $1.40 / $4.40, then GLM-5.1, GLM-5, GLM-4.7 and GLM-4.6. There is no GLM-5.5 row to price.
• The zai-org organisation on Hugging Face — its four newest text models, all created 25 August 2026, are GLM-5.3, GLM-5.3-BF16, GLM-5.3-Flash and GLM-5.3-Flash-BF16. Nothing numbered 5.5 is above them.
• A Hugging Face repository search for "GLM-5.5" returns two results and both are false positives: an EXL2 quantisation of a GLM-4-9B fine-tune and an EXL2 quantisation of GLM-4.5-Air, where "5.5" is a bits-per-weight figure in the filename rather than a model version. No repository anywhere on the Hub is a GLM-5.5.
• Artificial Analysis returns a live page for GLM-5.3; artificialanalysis.ai/models/glm-5-5 returns 404.
• The zai-org GitHub organisation's newest text-model repository is GLM-5, which now carries GLM-5.3, GLM-5.2, GLM-5.1 and GLM-5 together. No GLM-5.5 repository exists.
• Our own catalogue, read live today — see the route note further down.

None of that is a statement about obscurity, and it is worth saying why. GLM-5.3 has been public since August and leaves a checkable trail: a documentation page, a rate-card row, a Hugging Face repository with 1.7 million downloads at the time of writing, a model card with a declared licence. A shipped Z.ai flagship leaves that trail within days of the API opening. A number that has left none of it is not a quiet release. It is a number.

What the specification question actually resolves to

The brief for a reference page normally reads "give the specification, attributed." For GLM 5.5 the honest version of that is shorter than usual: every number is unstated, because there is no model card to state it. No parameter count, no context window, no modality list, no reasoning-effort parameter, no licence, no price. Where community threads print a figure for GLM 5.5, that figure is a projection about a model nobody has seen, and this page will not launder one into a specification.

What can be attributed is the line the number would join, and that is worth having because it tells a reader what "the next release" would plausibly be. GLM-5.3 is a 753-billion-parameter mixture-of-experts model — that total is read from the safetensors metadata on the lab's own zai-org/GLM-5.3 repository rather than from a vendor claim sheet — published in FP8 under the architecture name GlmMoeDsaForCausalLM. Z.ai's documentation gives it a 1,000,000-token context window with a maximum output of 128,000 tokens, text input and text output only, and no option to turn reasoning off: thinking is always enabled and depth is set with reasoning_effort at low, high or max, defaulting to max. GLM-5.3-Flash is a genuinely different model rather than a cut-down one — 320 billion total parameters with 18 billion activated, the first open frontier model in the line to combine sparse and linear attention, and natively multimodal with text, image, video and file input against text output.

Those two shapes are the two shapes a "5.4" or a "5.5" would plausibly take, and they are also the reason the naming argument inside Z.ai's own story is unresolved rather than merely early. GLM-5.3 was explicitly a post-training step: Z.ai's documentation states it uses the same base model as GLM-5.2 and that every gain comes from post-training. GLM-5.3-Flash was a newly trained base. So the line has both habits in it at once, and a reader watching for the next number cannot infer from either one alone whether it will be a point release on an existing base or a new one.

Licence and weights: read this one, do not infer it

This is the part of a release like this that a reader most often gets wrong by generalising from a sibling, and it is the reason this section exists even though GLM 5.5 has no licence at all. The two shipping models in the line are under different terms.

• GLM-5.3 — the Hugging Face repository declares license: other with license_name: glm-5.3. That is a custom licence written for this release, not MIT and not Apache, and its terms are not the permissive terms of the Flash model. If you plan to build a product on these weights, the licence text is a required read rather than a formality.
• GLM-5.3-Flash — MIT, declared on the lab's own repository. That is a real difference in what you are permitted to do with the weights, and for many teams it is the single strongest reason to pick Flash over the more capable flagship.
• What is downloadable — both are published in FP8 with separate BF16 repositories (GLM-5.3-BF16 and GLM-5.3-Flash-BF16) alongside them, so the format question has an answer that does not require conversion work before you start.
• What "will it fit" means — for GLM-5.3 the total is 753B parameters and your storage has to hold all of them; for GLM-5.3-Flash the 18B activated figure is the one that governs memory pressure and throughput at inference while 320B is the weight footprint. No throughput or VRAM figure for either model is stated here, because neither Z.ai nor the repositories publish one and a number this page cannot source is a number it will not print.

Running it: what a reader can actually obtain

Z.ai's own model card for GLM-5.3 names the supported serving frameworks, which is the closest thing to a deployment guide the lab publishes: SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth, with deployment on the Ascend NPU platform documented separately. GLM-5.3-Flash has been picked up by the quantisation community in the usual way, including MLX ports for Apple silicon, so a single-machine path exists for the smaller model as well as a datacentre one for the flagship. The weights for both landed on Hugging Face on 25 August 2026.

Practically, the split runs like this. If your work is text-only, repo-scale coding, long-horizon agent tasks or anything that lives or dies on a million-token context, GLM-5.3 is the model, and the custom licence is the constraint you accept in exchange. If your work needs images, screenshots, video or documents in the loop, or if you need MIT terms, GLM-5.3-Flash is the model, and the capability gap is the constraint you accept in exchange. Both are callable through Z.ai's own API and several third-party platforms, and GLM-5.3 is also on our catalogue, which is where the next section starts.

Availability: what we are and are not claiming

A reader who has seen Z.ai models on gateways will reasonably assume GLM 5.5 is reachable too, so this needs saying plainly rather than being left to inference. There is no GLM 5.5 route on OrcaRouter, and this page is not claiming one. Our live catalogue holds ten Z.ai identifiers — from z-ai/glm-4.5 and glm-4.5-air up through glm-4.6, glm-4.7, glm-5, glm-5.1, glm-5.2, glm-5.3, glm-5.3-flash and glm-5.3-flash-free — and no identifier matching glm-5.5 or glm5.5 under any spelling. That was checked against the live model list today, not assumed, and the same check returns nothing for glm-5.4 either.

What is reachable is the model that exists, and it is worth wiring now precisely because the next number is unknown. GLM-5.3 on OrcaRouter reads $1.26 per million input tokens and $3.96 per million output today, against Z.ai's own published rate card of $1.40 and $4.40 for the same model; GLM-5.3-Flash on the same key reads $0.15 and $0.50, matching Z.ai's card. We pass provider pricing through with no markup of our own, so a Z.ai price change shows up here the day it is published rather than at the next contract renewal — and if a GLM 5.4 or 5.5 does ship, the practical way to evaluate it without betting a production path on an unproven model is to point a slice of traffic at it behind automatic failover, with GLM-5.3 as the fallback leg. Same key, no second contract, no re-integration.

GLM 5.5 and GLM 5.4: the difference is which one is missing

The two numbers are adjacent, share a demand profile on our property — 1,415 impressions for 5.5 against 1,373 for 5.4 — and neither is a model, so the difference between them is not a specification difference. It is a difference in what observers expect and why. The 5.4 argument comes from cadence: Z.ai has shipped a flagship roughly every 54 to 70 days through the line, and the next slot on that arithmetic falls in a window that opened on 8 October. The 5.5 argument comes from the leak trail — the name surfaced as a pre-registration string in a third-party tool and in a leak post, and the anonymous model fingerprinted in August was guessed to be either 5.3-Vision or 5.5. Both are projections, neither is a vendor statement, and this page takes no position on which is more likely. The 5.4 half is tracked on our GLM 5.4 page; this page is the 5.5 half, and the two are deliberately separate.

The comparison set, as an index rather than a verdict

Nine comparison pages exist on this stem and they are the right place to go for a head-to-head. Listed without adjudicating, because in each of them one side has no specification and the useful content is the shipped model's numbers: claude-opus-5, claude-sonnet-5-5, gemini-4-argon, gpt-6, mistral-large-4-0, muse-wonder, reflection-beam, step-5-preview, deepseek-v4-pro, gemini-3-1-pro, grok-4-5, kimi-k3, minimax-m3 and qwen-3-8. Whichever one you open, the rule for reading it is the same one this page has applied: ask which side has a rate card.

What is not settled

Whether a GLM 5.5 exists at all, and whether the next Z.ai number is a point release on the GLM-5.3 base or a new base that skips the name entirely. Z.ai has confirmed neither, and there is no artefact on any of the nine surfaces above that a reader can check to settle it. The one thing genuinely settled is the specification question, in the negative: there is nothing to specify.

Which to pick, once there is a choice

If what you actually need is a shipped model this week, the decision does not involve 5.5. Pick GLM-5.3 when the work is text-only and long-horizon and you can live with a custom licence; pick GLM-5.3-Flash when you need images or documents in the loop, or when MIT terms decide it. If what brought you here was the name, the instruction is to wait for an identifier — a rate-card row, an API model id, or a repository under zai-org — and to judge the arrival on those rather than on a leaderboard screenshot. A version number with no model card is not a model, however many pages it already has.