GLM 5.4: What the Name Refers To, and What Z.ai Actually Ships Today

GLM 5.4 is a name, not a model. Search it today and you will find no model card, no API identifier, no Hugging Face checkpoint, no licence file, no rate-card row and no page on Z.ai's own documentation that uses the number. Nine surfaces were checked for this page on 13 October 2026 and every one of them came back negative. What does exist is the line it would join: GLM-5.3, Z.ai's text flagship, and GLM-5.3-Flash, the natively multimodal model that followed it in August 2026 — both released, both priced, both downloadable.

That gap is the whole reason this page exists. "GLM 5.4" draws a real, measured search on our property — 1,373 impressions over eight queries in the last twenty-eight days, 97% of them on the bare name — and until now nothing here answered it. What a reader typing it wants is simple: does this model exist, what is it, and if it does not, what do I run instead. The answers are no, nothing yet, and three models from the same lab that you can call this afternoon.

The nine surfaces, checked on 13 October 2026

A claim that a model does not exist decays the moment it is written, so here is the method and not just the conclusion. Every line below was re-run for this page, and every one is checkable by a reader in an afternoon.

• Z.ai's own release log — docs.z.ai/release-notes/new-released — leads with GLM-5.3-Flash, dated 26 August 2026. There is no entry above it.
• Z.ai's developer documentation index — docs.z.ai/llms.txt — lists four model pages: GLM-5.3-Flash/FlashX, GLM-5.3 and GLM-5.2. Nothing numbered 5.4.
• Z.ai's OpenAPI specification enumerates the callable identifiers: glm-5.3, glm-5.3-flash, glm-5.3-flashx, glm-5.2, glm-5.1, glm-5, and the 4.x line. No glm-5.4.
• The Chinese-language documentation set at docs.bigmodel.cn lists the same ceiling: GLM-5.3 and GLM-5.3-Flash.
• The zai-org organisation on Hugging Face — the lab's own weight repository — has GLM-5.3 and GLM-5.3-Flash as its newest text models, both created 25 August 2026. Nothing numbered 5.4 sits above them.
• A full-text search of Hugging Face for "glm-5.4" returns zero repositories, under any author.
• z.ai/blog/glm-5.4 returns 404, as does z.ai/blog/glm-5.5.
• Artificial Analysis has a live GLM-5.3 page; artificialanalysis.ai/models/glm-5-4 returns 404.
• OrcaRouter's own catalogue, read live today, serves ten Z.ai identifiers from z-ai/glm-4.5 to z-ai/glm-5.3-flash-free. There is no glm-5.4 among them, under any spelling.

None of that is a statement about obscurity. GLM-5.3 has been public since August and draws a large, checkable footprint — a model card, a rate-card row, a Hugging Face repository with millions of downloads. A shipped Z.ai flagship leaves that trail within days. Nothing numbered 5.4 has left one.

Where the number 5.4 came from

The name has a paper trail, and it is short enough to read in full.

The first appearance we recorded was on 25 September 2026, when the strings "GLM 5.5 Flash" and "GLM-5.4" turned up on a third-party coding tool's public data page. They were pre-registration placeholders — no users behind them, no vendor confirmation, no model attached — and when the same page was re-checked on 29 September, both strings were gone. A string inside somebody else's product is not a release. It is the kind of artefact that gets screenshotted into a leak story, and this one did.

The second appearance was in a leak post that put four unannounced names into circulation at once: a K4 model, GLM-5.5 Flash, GLM-5.4 and DeepSeek V4.1P. The post's own emphasis was on the K4 name and its activated-parameter arithmetic, not on either GLM number; both GLM names were cited as interesting nomenclature. No config, no checkpoint listing, no staging endpoint accompanied them.

The third source is not a sighting at all but arithmetic. Z.ai has shipped a flagship roughly every 54 to 70 days — GLM-5 on 12 February 2026, GLM-5.1 on 7 April, GLM-5.2 on 16 June, GLM-5.3 on 18 August — and analysts reading that cadence have put the next one in a window running from 8 October to 9 November 2026, with most expecting it to be numbered GLM-5.4 rather than GLM-5.5. Today is 13 October and that window is open, which is the most concrete thing anyone can currently say about the name. It is a projection, not a vendor statement.

Why analysts expect 5.4 while Z.ai named 6.0

These two facts are usually reported separately and they are much more useful together, because between them they explain why "GLM 5.4" is simultaneously the most likely next number and a number the lab's own leadership has never spoken.

The expectation is structural. Every flagship in this line has been a point release on the previous one: 5, 5.1, 5.2, 5.3. Nothing in the family has ever jumped a number, and GLM-5.3's own documentation is explicit that it is not a new base model — it uses the same base as GLM-5.2 and takes all of its gains from post-training. A lab that improves a model by post-training it and bumps the point release is a lab whose next text flagship, on pattern, is called GLM-5.4.

The countervailing fact is what the founder said. On Z.ai's first interim earnings call as a listed company, on 31 August 2026, Tang Jie named the next-generation base model on the record as GLM-6.0, positioned around a training route the company calls "Full Self-Training." That is a new generation rather than a point release, and it skips 5.4 and 5.5 entirely. Note what it does not do: a named future generation does not cancel a smaller step in between, and Z.ai has shipped point releases under an announced bigger roadmap before. It does mean that when GLM-5.4 does or does not arrive, the number will not be the interesting part — whether it is a post-training step on the GLM-5.3 base, like GLM-5.3 was on GLM-5.2, or a genuinely new base will be.

What Z.ai actually ships today

This is the part a reader searching "GLM 5.4" can act on, and it is worth reading as a line rather than as three isolated product pages. The dates are Z.ai's own release log; the sizes are from the lab's published weights.

• GLM-5 — 12 February 2026. The first GLM-5, built for complex systems engineering and long-horizon agent tasks, with sparse attention for token efficiency.
• GLM-5.1 — 7 April 2026. Aimed at long-horizon work, able to run solo for up to eight hours in a single session.
• GLM-5.2 — 16 June 2026. The 1M-context flagship, pitched at reducing context drift and goal forgetting on long tasks.
• GLM-5.3 — API open 18 August 2026, weights published to Hugging Face on 25 August. The current text flagship.
• GLM-5.3-Flash — 26 August 2026. A different, smaller, natively multimodal model on a newly trained base.
• GLM-5.3-FlashX — the same Flash weights served at a higher price tier: $0.37 input / $1.25 output per million tokens against Flash's $0.15 / $0.50.

The specification of GLM-5.3, which is what a "GLM 5.4" page would most likely be describing the successor to, reads like this. It 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, not from a vendor claim sheet. It carries a 1,000,000-token context window with a maximum output of 128,000 tokens. It is text-only: text in, text out. Reasoning is not optional — GLM-5.3 always operates with thinking enabled, and the depth is set with a reasoning_effort parameter taking low, high or max, defaulting to max. Z.ai's migration guide recommends keeping max for coding and complex reasoning, and warns that requests still sending thinking.type: "disabled" will fail against the new identifier.

GLM-5.3-Flash is the more interesting of the two for anyone building on a budget, and it is a different model rather than a cut-down GLM-5.3. Z.ai's release notes describe 320 billion total parameters with 18 billion activated, a hybrid of linear and sparse attention that cuts compute and KV-cache requirements, and native visual capability — it observes interfaces and rendering results, which closes the loop on code, browser and GUI work. It is also the model Z.ai chose as the foundation for GLM-Legal, the legal agent it published on 12 October 2026, which is a reasonable signal of where it sits in the line.

Licence and weights: what is actually downloadable

For a reader who arrived asking which files to pull and whether they will fit, the distinction between these two models is the one that matters, and it is a licence distinction before it is a size one.

• GLM-5.3 — the licence is not MIT and not Apache. The Hugging Face repository declares license: other with license_name: glm-5.3, a custom licence written for this release. Read it before you build a product on the weights; the terms are Z.ai's own and they are not the permissive terms of the Flash model.
• GLM-5.3-Flash — MIT, on the lab's own repository tags. That is a genuine difference in what you are allowed to do with the weights, and it is the single strongest reason a builder picks Flash over the flagship despite the capability gap.
• Serving them yourself — Z.ai's own model card names the supported frameworks: SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth, with Ascend NPU support documented separately. The flagship's repository ships FP8 weights; the community has already produced a wide range of quantisations and MLX ports for Flash, so both a datacentre deployment and a single-machine one are covered.
• What "fit on my machine" means here — treat 753B total parameters as the headline for GLM-5.3 and 320B total / 18B active for Flash. The activated-parameter figure is the one that governs throughput and memory pressure at inference; the total is what your storage has to hold.

What you can call today, and the route we are not claiming

To be plain about the question a gateway-shaped answer invites: there is no GLM 5.4 route on OrcaRouter, and this page is not claiming one. We hold no model identifier matching glm-5.4 or glm5.4 under any spelling, and that was verified against our live catalogue on 13 October 2026 rather than assumed. If you have seen other GLM identifiers on a gateway and are assuming this release is reachable, it is not.

What is reachable is the model that exists. GLM-5.3 is on our catalogue, and GLM-5.3 is on our catalogue at a rate read live today: $1.26 input and $3.96 output per million tokens, with cache reads at $0.234. Z.ai's own published rate card lists GLM-5.3 at $1.40 / $4.40 with cached input at $0.26 — that is the vendor's number, and it is the one to compare against if you are pricing this out yourself. GLM-5.3-Flash sits beside it at $0.15 / $0.50, matching Z.ai's card. Z.ai also serves both through its own API, and both are available through several third-party platforms.

GLM 5.4 and GLM 5.5, side by side

The two names share an origin and almost nothing else, and the prospect of one arriving should not be read as the other being replaced. Both were pre-registration strings inside the same third-party tool in late September 2026 and both vanished on re-check days later. From there they diverge: 5.4 is the number the cadence argues for and the one most analysts expect the next point release to carry, while 5.5 is the number that got skipped outright when Z.ai's founder named GLM-6.0 as the next-generation base on 31 August. If you are trying to work out which of the two a future release will be called, the honest answer is that the vendor has confirmed neither and the naming direction is the one thing in this line that has historically been read upward by observers rather than announced downward by the lab. Our glm-5-5-release-date page tracks that half of the question; this page tracks the other, and the demand is nearly identical on both sides — the two names measure 1,373 and 1,415 impressions respectively on our property, and neither is a model anyone can serve.

What is not settled

Everything about GLM-5.4 except the name. There is no release date, no parameter count, no context window, no licence, no price and no route, and any page that prints those figures is inventing them. The one live question is whether the next Z.ai release is a point step on the GLM-5.3 base or the new-generation GLM-6.0 base the founder named — the two are different products with different implications for anyone planning an upgrade, and no source settles it today.

What would have to appear before this page changes

Five artefacts, in the order they normally arrive, and the first is the cheapest for a reader to watch themselves.

• A fourth card on Z.ai's model-api documentation, or a new identifier in its OpenAPI specification. This is the first thing that moves and it costs nothing to check.
• A rate-card row on docs.z.ai/guides/overview/pricing carrying the new identifier and a per-million-token price on both lines.
• A Hugging Face repository under zai-org with the weights, a declared licence at the top of its model card, and a parameter count in the safetensors metadata.
• An entry in the release log above the 26 August 2026 GLM-5.3-Flash line.
• An identifier that resolves — the test this page used on our own catalogue, and the one any reader can repeat against whichever platform they use.

Until one of those lands, the model a reader can actually run is the one from August. GLM-5.3 handles text-only long-horizon work with a 1M-token window under a custom licence; GLM-5.3-Flash handles the same class of tasks with images and GUIs in the loop, on 18 billion activated parameters, under MIT. Both are callable now, and picking between them is a licence and modality decision rather than a capability gamble. GLM 5.4 is a number with a projection attached to it, and a projection is not something you can put in a production path.