GLM-5.3's License Split: MIT for Flash, Not for the Flagship

Z.ai shipped two models named GLM-5.3 within the same week, and most of the early coverage treated them as one release with one license. They are not. The smaller of the two is plain MIT. The larger one, published two days later, carries a bespoke license with a clause aimed squarely at a handful of the largest companies on earth.

If you are the kind of person who reads “open weights” and stops there, this is the release that shows why that habit gets you in trouble.

Two Releases, Two Days Apart

GLM-5.3-Flash went up on Hugging Face on 26 August 2026: a 320 billion parameter mixture-of-experts model with roughly 18 billion active per token, a hybrid architecture combining sparse and linear attention, a context window past 1 million tokens, and native multimodal input across text, image, and video. Z.ai’s model repository lists the weights under a plain, unmodified MIT license.

Two days later, on 28 August, Z.ai published the flagship GLM-5.3: roughly 753 billion total parameters with about 40 billion active, built for longer-horizon agentic and software engineering work rather than everyday tasks. This is the model most of the “GLM-5.3 is open weights” headlines were actually about. Its Hugging Face model card tags the license as a custom “glm-5.3” license, not MIT and not an OSI-approved license at all.

That is the detail worth separating from the specs, because “GLM-5.3” now names two products with two different legal answers to “can I build a business on this.”

What The Bespoke License Actually Requires

For nearly every reader of this post, nothing changes. You can download the flagship weights, run them, fine-tune them, and ship a commercial product built on them, the same as with a permissive license.

The clause that makes this a different license rather than MIT with a new name attached applies to one narrow case: a Model-as-a-Service business whose revenue exceeds 10 billion dollars over a trailing twelve-month period has to pass a security review determined by Z.ai before continuing commercial use. Coverage of the license text quotes the operative line as leaving the process entirely to the vendor’s discretion, with no published criteria, timeline, or appeal route spelled out. GLM-5.2, the prior flagship, shipped under plain MIT with no such clause, which is why the change registered as news rather than a formality.

That is a meaningful shift in what “open weights” is committing to, even though it changes nothing for almost everyone who will actually download the model. The broader vocabulary for reading a clause like this against Apache, MIT, or a community license is covered in open weights model licenses: Apache, MIT, and Llama.

Who The Clause Is Actually Aimed At

A 10 billion dollar trailing revenue threshold, scoped to companies reselling inference or fine-tuning access to third parties, describes a very short list: the hyperscale cloud platforms and the largest API resellers. It does not describe a startup, a mid-size company running the model internally, or an individual self-hosting it for personal use.

Read uncharitably, a discretionary security review with no published criteria is a lever a vendor could use against exactly the companies it is aimed at, without much recourse. Read more generously, it is a narrow attempt to keep the biggest platforms from repackaging the weights without any relationship to Z.ai at all. Neither reading changes the practical answer for a normal reader: this is not the clause that affects you, in the same way the Model-as-a-Service condition in Qwen’s community license does not affect most people running that model either.

Neither Model Runs On Your Laptop

Set the license question aside and the hardware question answers itself the same way for both releases. 320 billion total parameters and 753 billion total parameters both require every expert resident in memory for the mixture-of-experts routing to work, regardless of how few activate per token. Active parameters set compute cost; total parameters set the memory floor, and that floor sits well into multi-GPU server territory for both models. The distinction is covered in more depth in mixture of experts explained for writers and can you run a trillion parameter model locally.

If your interest in either release is a private, offline writing tool, neither GLM-5.3 nor GLM-5.3-Flash is the model for that job. The size class that actually runs on a laptop is one to four billion parameters, dense, quantized, which is a different product wearing the same version number.

What This Changes For Your Status Update

The failure mode this release invites is the same one every dual-release invites: the two products blur into one sentence by the time the news reaches a decision maker.

Before:

Z.ai released GLM-5.3 as open weights, so licensing shouldn’t be a blocker if we want to try it.

After:

Z.ai released two models this week. GLM-5.3-Flash (26 August) is MIT licensed with no restrictions. The larger flagship GLM-5.3 (28 August) uses a different license that requires a Z.ai security review only for Model-as-a-Service businesses over 10 billion dollars in trailing revenue, which does not apply to us. Worth confirming which model any benchmark we look at actually used.

The second version survives a legal or procurement read without someone having to go verify which GLM-5.3 the first version meant.

A Wrivio Context for AI model status updates could say:

Rewrite this as an internal note about a new AI model release. Neutral, factual register. Keep every date, parameter count, and license term exactly as written. If the update conflates two separate model releases, separate them rather than smoothing them into one claim. Do not add a recommendation to adopt the model unless one is already present in the original.

Press Ctrl+Shift+Space, paste the draft, and check the diff. A model that holds the distinction between the two releases is doing the actual work; one that collapses them back into “GLM-5.3 shipped open weights” has erased the one detail that mattered.

Common Questions

Is GLM-5.3 open source?

GLM-5.3-Flash is MIT licensed, which does meet the open source definition. The larger flagship GLM-5.3 uses a custom license with a revenue-based restriction, which is open weights but not open source in the strict, OSI sense.

Does the GLM-5.3 license stop most companies from using it commercially?

No. General commercial use, including inside a product you sell, is unrestricted. The security review clause applies only to Model-as-a-Service businesses clearing 10 billion dollars in trailing twelve-month revenue.

Can I run either model on a consumer laptop?

No. Both are mixture-of-experts models with hundreds of billions of total parameters, which have to be resident in memory regardless of how few activate per token, putting both well past consumer hardware.

Why did Z.ai change the license between GLM-5.2 and the GLM-5.3 flagship?

Z.ai has not published a plain-language rationale. The practical effect described in early coverage is a narrow clause targeting the largest resellers of inference, not a broad restriction on ordinary use.

Is GLM-5.3-Flash the same model as the GLM-5.3 flagship, just smaller?

They share a version number and an architecture family, but they are separate releases with separate specifications, separate release dates, and separate licenses. Confirm which one a given benchmark or announcement is actually describing.

Download Wrivio for Windows to keep drafting on a small, Apache 2.0 local model while the largest open-weight releases sort out what their licenses actually say.