Is GLM Open Source? Licenses for Every GLM Model (MIT vs GLM-5.3 License) and Commercial Use

Is GLM open source? Mostly, yes. Z.ai (formerly Zhipu AI) publishes the weights of almost every major GLM model on Hugging Face and ModelScope. GLM-5.3-Flash, GLM-5.2, GLM-5.1, GLM-5, GLM-4.7, GLM-4.7-Flash, GLM-4.6, GLM-4.6V, GLM-4.5, GLM-4.5-Air and GLM-Image all use the MIT license, which allows commercial use, modification and redistribution with almost no strings attached.

There are two exceptions to know about. The flagship GLM-5.3 has open weights under a custom GLM-5.3 License: MIT-style permissions plus one extra condition that only affects very large “Model as a Service” providers. And some GLM products are API-only, with no weights at all: GLM-5-Turbo and the faster X variants (GLM-5.3-FlashX, GLM-4.7-FlashX, GLM-4.5-X, GLM-4.5-AirX). This guide gives the license for every model, explains both licenses in plain language and walks through the common commercial-use cases. It is a practical summary, not legal advice.

GLM license for every model

The table covers every current GLM model with its weights status, license and official repository. Licenses come from the Hugging Face model metadata and model cards under the zai-org organization; ModelScope mirrors the same repos under ZhipuAI.

The pattern is easy to remember: base models are open, speed tiers and specialist variants are not. The X and FlashX models are the same families served faster on Z.ai’s infrastructure, and GLM-5-Turbo is a GLM-5 variant trained for OpenClaw agent workflows that Z.ai keeps behind its API. If a model has a page on this site, the “Downloading the weights” section there shows the exact repos and hardware notes.

What the MIT license lets you do with GLM

MIT is one of the most permissive licenses in software. For the GLM models released under it, that means:

• Use it for anything, including commercial products, internal tools and paid services. No revenue caps, user caps or field-of-use limits.
• Modify it. Fine-tune, quantize, distill, merge or prune the weights, and keep your changes private if you want.
• Redistribute it, original or modified, for free or for money, and sublicense it under your own terms.
• Host it for others, including as a paid API.

There is one obligation: keep the copyright notice and the license text with all copies or substantial portions of the weights and code you redistribute. And there is one disclaimer: the software comes “as is”, with no warranty, so if the model gives a wrong answer in your product, the license gives you no claim against Z.ai.

Z.ai leans on this openness in its own messaging. The GLM-5.2 announcement describes the release as “Pure Open: an MIT open-source license — no regional limits”. The MIT license text itself contains no acceptable-use list, no user-count threshold and no naming rule for derivatives, so there is nothing extra to track once the notice is in place.

The GLM-5.3 License explained

Every GLM flagship from GLM-4.5 to GLM-5.2 shipped under MIT; GLM-5.3 is the first to break that pattern. Z.ai launched it on August 14, 2026, held the weights back for about two weeks for safety evaluation, and published them on August 25 under the new GLM-5.3 License. The license has three clauses.

Clause 1: MIT-style permissions

The grant is broad and free of charge. It covers the model weights, parameters, configuration files, inference and training code and documentation, and lets any person or entity use, copy, modify, merge, publish, distribute, sublicense and sell copies, and “run, deploy, fine-tune, or otherwise modify the Software and create derivative works from it”. As with MIT, you must include the copyright and permission notice in copies, and your use must comply with applicable laws and regulations.

Clause 2: the Model as a Service condition

This is the only real difference from MIT. It starts with a definition:

“Model as a Service” means giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.

GLM-5.3 License, clause 2

The condition then applies only if both of these are true: the licensee or any of its affiliates operates a Model as a Service business, and the combined revenue of the licensee and its affiliates exceeds $10 billion (or the equivalent in other currencies) over any consecutive 12 months. In that case, the license says, “the Licensee must pass Z.AI’s security review before using the Software or its derivative works for any commercial purpose.” Z.ai sets the scope and method of that review “reasonably”.

Three practical readings follow from the text:

• For almost every company, individual and research group, GLM-5.3 behaves exactly like an MIT model. The revenue threshold is measured across the whole corporate group, and very few groups exceed $10 billion a year.
• Even a group above the threshold is only caught if it runs a Model as a Service business, meaning it gives third parties meaningful control over inputs, parameters or training data (a general-purpose inference or fine-tuning API, for example).
• Even then, the consequence is a security review before commercial use, not a ban or a fee. Non-commercial use is not affected by clause 2.

Clause 3: no warranty

Standard MIT-style disclaimer: the software and its outputs come “as is”, and Z.ai, its affiliates and copyright holders are not liable for claims arising from its use. Questions about the license go to glmlicense@z.ai. The full text is in the GLM-5.3 repository on Hugging Face.

GLM open source vs open weights

You will see GLM described as both “open source” and “open weights”. Both are used, and the difference matters if you are writing a policy or a procurement document.

• Open weights means the trained parameters are downloadable and you can run, fine-tune and redistribute them. Every GLM model marked “Open” in the table meets this definition.
• Open source, in the strict sense many software people use, also implies you could reproduce the model: training data, full training code and recipes. The GLM releases center on weights, configuration, chat templates and inference support, backed by technical reports (arXiv 2508.06471 for GLM-4.5, arXiv 2602.15763 for GLM-5) and Z.ai’s open-source RL framework slime.

Z.ai itself uses “open source” freely: its release notes call GLM-4.5V “a 100B-scale open-source vision reasoning model”, and the GLM-5.3 announcement has an “Open Source” heading for the weight release. The most precise wording is: GLM models are open-weight models, most of them under the MIT license. Whatever term you use, the license text is what governs what you may do.

For most readers the label matters less than four practical questions: can you download the model, run it on your own hardware, change it, and ship it in a product? For every MIT GLM model the answer to all four is yes, with the notice as the only duty. For GLM-5.3 the answer is also yes for almost everyone, with the security-review condition reserved for the very largest hosting businesses.

GLM commercial use: common scenarios

Here is how the licenses apply to the situations people actually ask about. Again, this is not legal advice; for a large deployment, have counsel read the license file in the exact repo you use.

You self-host GLM inside your own product

A support chatbot, a coding assistant in your IDE, a document summarizer in your SaaS app. With any MIT model: allowed, no conditions beyond keeping the notice with redistributed weights. With GLM-5.3: allowed too, and clause 2 explicitly excludes “end-user products with model capabilities solely embedded within specific features or harnesses” from the Model as a Service definition.

You fine-tune GLM and sell the result

Allowed under MIT and under the GLM-5.3 License, which names fine-tuning and derivative works in its grant. Keep the original copyright and license notice in your distribution. Z.ai lists slime (v0.3.0+) and ms-swift (v4.4.0+) as supported fine-tuning frameworks for the GLM-5 series.

You sell GLM inference as an API

With MIT models: allowed. That is how third-party hosts can serve GLM models; OpenRouter, for example, routes GLM-5.2 requests across several providers. With GLM-5.3: this is the Model as a Service case. You may do it, and only if your group’s revenue is above the $10 billion threshold do you need Z.ai’s security review before commercial use.

You route requests to GLM hosted by someone else

A gateway or router that forwards requests to Z.ai’s API or another host is not running the weights at all, and the GLM-5.3 License excludes “mere relaying of requests to models hosted by others” from Model as a Service anyway. What governs you here is the terms of the service you call.

You use GLM for research, teaching or personal projects

Allowed under every open GLM license, with nothing to report or request. You can publish benchmark results, fine-tuned checkpoints and papers built on the weights; if you redistribute the weights themselves, include the notice. The GLM-5.3 License’s extra condition only concerns commercial use by very large Model as a Service groups, so it never applies to non-commercial research. If you publish work built on GLM-5 series models, Z.ai asks you to cite the GLM-5 technical report (arXiv 2602.15763).

You use the Z.ai API in a commercial app

Weight licenses do not apply to API use; Z.ai’s terms of use do. They grant a non-exclusive right to integrate the API into your applications for your end users, say Z.ai will not claim ownership of your inputs and your end users’ content, and, as between you and Z.ai, leave you the rights in the outputs generated for you. The terms also require AI-generated outputs to be marked as AI-generated. API-only models such as GLM-5-Turbo and the FlashX variants can only be used this way. Start with the GLM API quickstart and check costs on the GLM pricing page.

Is Z.ai GLM-5.1 open source?

Yes. GLM-5.1, released on April 7, 2026, is fully open-weight under the MIT license. The weights are on Hugging Face as zai-org/GLM-5.1 (BF16) and zai-org/GLM-5.1-FP8, and on ModelScope as ZhipuAI/GLM-5.1. It is a 744B-parameter mixture-of-experts model with 40B active parameters, a 200K context and 128K maximum output, built for long-horizon agent work.

Because GLM-5.3 reuses the GLM-5.2 base and changed license, some readers assume the whole GLM-5 line moved away from MIT. It did not: GLM-5, GLM-5.1 and GLM-5.2 stay MIT, and the newer GLM-5.3-Flash is MIT as well. Only GLM-5.3 uses the GLM-5.3 License. More about the model on the GLM-5.1 page.

Where to download GLM weights

Get weights only from the official organizations: huggingface.co/zai-org and ModelScope under ZhipuAI. The GLM-5 GitHub README has the full download table for the GLM-5 series and lists serving options:

• Serving: SGLang, vLLM, Transformers, KTransformers and Unsloth for the GLM-5 series; TokenSpeed is also listed for GLM-5.3-Flash; vLLM-Ascend, xLLM and SGLang for Ascend NPU deployments.
• Precision: most GLM-5 series models ship in BF16 and FP8. GLM-5.3 and GLM-5.3-Flash use FP8 as the default repo and add a -BF16 repo; GLM-5.2, GLM-5.1 and GLM-5 use BF16 as default and add -FP8.

Plan for the size. As a rough estimate for the weights alone (parameter count × bytes per parameter, ignoring KV cache and activations): a 744B model is about 744 GB in FP8 and about 1.5 TB in BF16; GLM-5.3-Flash at 320B is about 320 GB in FP8; GLM-4.7-Flash at 31B is about 62 GB in BF16. That is why GLM-4.7-Flash, a 30B-A3B model and the most-downloaded GLM repo with more than 1.8 million downloads a month, is where most local users start.

Open GLM models by size

An open license is only useful if you can run the model. This table lists the open GLM models by parameter count so you can match a license to your hardware. All the large ones are mixture-of-experts models: only the “active” parameters run for each token, but all parameters must sit in memory.

The smallest models are where “open” pays off most for individuals and small teams. GLM-4.7-Flash and GLM-4.5-Air run on far less hardware than the 744B flagships, and both are MIT, so you can ship them inside a commercial product without any review. For the 744B models, most teams either rent large GPU clusters or use the hosted API; the GLM-5.2 page covers what running the big models locally involves.

How to check a GLM license yourself

Licenses can change between releases, as GLM-5.3 showed. Before you ship a model, confirm the license in the exact repository you download from:

1. Open the model page under huggingface.co/zai-org and look at the license tag in the model card header. MIT models show license: mit; GLM-5.3 shows license: other with license_name: glm-5.3.
2. Open the LICENSE file in the repo’s file list and read it in full. That file, not a blog post or a third-party summary (including this one), is what binds you.
3. If you use a quantized or fine-tuned copy from another uploader, check that it kept the original license and notice. Derivatives of MIT models must carry the notice; derivatives of GLM-5.3 carry the GLM-5.3 License conditions.
4. For scripted checks, query the Hugging Face API, which returns the license among the repo tags.

curl -s https://huggingface.co/api/models/zai-org/GLM-5.3 | python3 -c \
  "import json,sys; d=json.load(sys.stdin); print([t for t in d['tags'] if t.startswith('license')])"
# ['license:other']

curl -s https://huggingface.co/api/models/zai-org/GLM-5.2 | python3 -c \
  "import json,sys; d=json.load(sys.stdin); print([t for t in d['tags'] if t.startswith('license')])"
# ['license:mit']

Z.ai’s open-source track record

Open releases are not new for Z.ai. According to its company timeline, it open-sourced the 100B-parameter GLM-130B in August 2022; the open ChatGLM-6B from 2023 has been downloaded more than 20 million times; and GLM-4-9B and GLM-4V-9B were open-sourced in June 2024. Check each older repo’s own license file before reusing those models, since their terms predate the MIT releases of the GLM-4.5 era. The company’s history is covered on the Zhipu AI page, and release dates for every model are on the GLM release timeline.

Is GLM open source? FAQ

Is GLM open source?

Mostly. The main GLM models have open weights, and nearly all of them use the MIT license. GLM-5.3 uses the custom GLM-5.3 License, and a few variants (GLM-5-Turbo and the X/FlashX speed tiers) are API-only.

Is GLM-5.3 open source?

Its weights are open (published on Hugging Face on August 25, 2026) under the GLM-5.3 License. It grants MIT-style rights, with one added condition: Model as a Service businesses whose group revenue exceeds $10 billion over any 12 consecutive months must pass Z.ai’s security review before commercial use.

Can I use GLM commercially?

Yes. MIT models can be used commercially without conditions beyond keeping the license notice. GLM-5.3 can be used commercially by everyone except the rare very large Model as a Service providers, who need a security review first. API use is governed by Z.ai’s terms of use.

What license does GLM-4.5 use?

MIT, for both GLM-4.5 and GLM-4.5-Air. The same is true for GLM-4.6, GLM-4.7 and GLM-4.7-Flash.

Is GLM-5-Turbo open source?

No. Z.ai has not published GLM-5-Turbo weights. You can use it through Z.ai’s services and through OpenRouter (z-ai/glm-5-turbo).

Do I have to credit Z.ai when I use GLM?

If you redistribute the weights or code, you must include the copyright and license notice. If you only run the model to power your product, the MIT license does not require on-screen credit, though Z.ai’s API terms require you to mark AI-generated outputs as AI-generated.

Is the Z.ai API open source?

No. The API is a paid hosted service governed by Z.ai’s terms of use. What is open is the model weights: you can run the same MIT models yourself instead of calling the API, or use the free API models such as GLM-4.7-Flash.

Can I run GLM offline?

Yes, any model with open weights. Download it from huggingface.co/zai-org and serve it with vLLM, SGLang or Transformers. The small GLM-4.7-Flash is the practical starting point; the 744B models need data-center hardware.

Is GLM-5.3-Flash open source?

Yes, under MIT, with FP8 and BF16 repos on Hugging Face. Its faster sibling GLM-5.3-FlashX is API-only.

You do not need to download anything to try these models: the open GLM-5.3-Flash, GLM-5.2 and GLM-4.7-Flash are all in the free GLM chat, or jump straight to chat with GLM-5.3-Flash. For a side-by-side of every model, see GLM models compared.