What is GLM, and what was Ox Alpha

A model turned up with no name on it

In the second half of August 2026, a model called ox-alpha started answering questions on OpenRouter and inside the OpenCode terminal tool. It arrived with no vendor attached, no model card, no announcement and no benchmark chart. People used it anyway, because it was in the list and because the answers were good, and within days it was the model developers were arguing about.

On 26 August, Z.ai ended the guessing itself. Its launch post for GLM-5.3-Flash says, three paragraphs in: "Before release, we tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback." So the attribution to Z.ai came from Z.ai, published on the day the weights went up. Nobody outside the company had to prove it.

Because the story travels faster than its sourcing, here is the chain we actually opened. The vendor claim is on Z.ai's own blog post at z.ai/blog/glm-5.3-flash, which carries the date 2026-08-26 in the page itself. The weights are in the Hugging Face repository zai-org/GLM-5.3-Flash, whose commit history shows the repo created on 25 August and the model files uploaded across 26 August, matching the blog. The distribution side lines up too: OpenRouter's public model API gives z-ai/glm-5.3-flash a creation timestamp of 26 August 2026 at 13:59 UTC. Three separate records, and one date across all of them.

One part of the story we could not pin down, so we are telling you rather than filling the gap. Nobody has published the date ox-alpha first appeared, and Z.ai's post does not give a duration for the experiment either. The closest dated account is the coverage on the day of the reveal, which says developers had been guessing "for the better part of a week" (OfficeChai, 26 August). The ox-alpha listing has since been pulled from OpenRouter, so there is no public record left to check that against. If you see a precise duration quoted anywhere, it is worth asking where the number came from.

The reveal is the part that got written up. The week before it is the part worth keeping. For that week, a large number of people formed a working opinion about a model while knowing nothing about who built it, how big it was, what it scored, or where it came from. They had the answers and nothing else. That is about as close to a blind test as this industry has managed, and it happened by accident.

So what is GLM?

GLM is the model line published by Z.ai, a Chinese lab that also goes by Zhipu AI and uses both names on its Hugging Face organisation page. The line goes back to the ChatGLM releases and has been carried forward through GLM-4 and GLM-5, and if you have run into it as a consumer it was most likely inside a coding tool, because that is where it has been strongest.

GLM-5.3-Flash is the model Ox Alpha turned out to be. Its own model card opens by calling it "the first natively multimodal model in the GLM-5 series", with 320 billion total parameters and 18 billion active. That second number is the one that matters for how it feels to use. It is a mixture-of-experts design, so the full 320 billion parameters sit in memory but only a small slice of them is engaged for any given token, roughly the way a large organisation keeps hundreds of specialists on staff and pulls in three of them for a specific question. Z.ai's own description credits a hybrid of sparse and linear attention for holding long documents together without the usual cost of doing so.

The launch came with a benchmark chart, as launches do. Some of the benchmarks on it are Z.ai's own, including a test named Z.ai Code Bench, which makes them vendor-reported by definition. Independent trackers exist and we opened one while writing this. We have deliberately left its numbers out, because those pages serve a live figure that recalculates as models are added and re-run, and a score printed here today could be out of date within weeks while still looking authoritative. A benchmark number without a date attached is worse than no number at all.

Which is exactly why the anonymous week is the more durable evidence. Nobody could tune a submission to it, nobody could pick a favourable harness, and nobody was grading the brand, because there was no brand to grade. When the name finally landed on the model, the opinions people had already formed did not move. That is a rare thing to be able to say about any model release.

Two licences, two days apart

Two days after the reveal, Z.ai published a second set of weights, and the difference between the two documents nobody reads is the part worth a minute of your time.

GLM-5.3-Flash shipped on 26 August under the plain MIT licence, stated as license: mit in the repository's own metadata. Two days later the larger GLM-5.3 flagship arrived in a separate repository, zai-org/GLM-5.3, first commit titled "Initial commit 0828" on 27 August with the files landing on the 28th. Its metadata reads license: other, license_name: glm-5.3, which is a bespoke document written for this one model rather than anything off the shelf.

The GLM-5.3 licence file reads like MIT for its first paragraph and then adds a clause. Quoting it directly: "If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 10 billion US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose." The same clause adds that "The scope and method of the security review shall be reasonably determined by Z.AI." There is no published standard for what passing looks like and no appeal named in the text.

No reader of this article will ever trip that threshold. MultiChats is not remotely close to it either. The reason it belongs in this article is that both releases were reported with the same word, open, applied without qualification, from the same company inside the same week. The only way to tell them apart was to open two LICENSE files, which takes about a minute and almost nobody does.

In practical terms the two documents grant almost the same thing. MIT lets anyone use, modify, sell and redistribute the weights with the copyright notice kept intact, and the GLM-5.3 licence copies that language nearly word for word before the extra clause. The gap opens only for very large companies reselling inference. For everyone else the two are functionally identical, which is precisely why the difference slid past so much of the coverage.

That is a habit worth borrowing beyond this one story. The terms attached to a model are among the few things about it that are fixed, published and checkable. Scores get quoted constantly and licences get quoted almost never, and the licence is the one that will still be accurate next month.

Which GLM models you can actually open in MultiChats

We carry four, out of 59 models in the app as of 19 October 2026. The table below comes from the model registry rather than from any announcement.

"Always on" is literal. Z.ai's API does not accept a request to turn thinking off on these models, so our entries mark reasoning as non-disableable and you will see the thinking pass on every reply. Some people find that useful and some find it noisy, and it is worth knowing before you pick one for quick back-and-forth work.

The honest limits, in one place. None of the four will accept a PDF in MultiChats; the attach button will not let you. None of them has our web search tool wired up, so if a question needs live information you want a different model for that turn. GLM 5.3 Flash takes an image and the other three do not. Upstream, GLM 5.3 Flash can also take video, and we do not pass video to it, so do not build a habit around that. GLM 5.3 Flash carries the Recommended badge in our model picker. Since 18 September, each GLM 5.2 reply counts as two messages against your allowance, which the picker shows as "2× credits".

On context, both free entries advertise very large windows, and we are not going to quote the figure, for the same reason we skipped the benchmark. What a model advertises and the length at which its answers stay reliable are two different quantities, and the second one moves as we tune it. What you get instead is a meter in the composer that tells you, for the model you have selected, when you are approaching the point where answer quality tends to fall off. If that distinction is new to you, we wrote it up properly in what a context window actually is.

One more thing that is specific to models of this origin and that we can show in code. GLM requests from MultiChats are pinned to a vetted list of US and EU hosts and can never reach Z.ai's own endpoint. That is a control we chose, and one we can point at in the code. On the wider question of training: we do not train on your conversations, some providers might depending on which model you pick, and where we know of a difference, that model's info card carries a notice. The GLM cards carry none, because the routing pin keeps these requests off hosts that collect data.

Run your own version of the blind week

The useful lesson from Ox Alpha is a method, and it costs you almost nothing to copy.

Take a task where you already know what a good answer looks like, because you have done it before and can grade it honestly. A difficult email you have already sent, say, or a long document you have already summarised for someone. Then run it past two or three models over a few days and judge only what comes back, ignoring the name at the top, the announcement you read last week, and the leaderboard position you half remember. In MultiChats you switch models on your next message inside the same thread, and the conversation carries over, so this takes minutes rather than a project.

What you learn from that is not transferable to anyone else, and that is the point of it. The published scores measure a fixed set of tasks that are almost certainly not yours. Ox Alpha got a week of exactly this kind of judgement from thousands of people at once, purely because Z.ai withheld the label, and the verdict survived the label arriving. Your own version has a sample size of one, and it is about your work, which makes it the more useful of the two.

If you want a starting point, the GLM line built its reputation on code and on long tasks that take many steps to finish, and GLM 5.3 Flash is the first in the family here that will also look at a picture. Give it a long email thread you need reduced to a list of decisions, or a screenshot of a chart you want explained back to you in plain words. Then hand the identical job to a model from a different lab and see which reply you would actually send. That comparison is the whole exercise and it fits in an afternoon.

GLM 5.3 Flash and GLM 5.2 both sit on the free plan, which is the easiest place to run a blind week of your own. What each plan includes is on the home page for whenever that question becomes the relevant one.