Open weights
GLM-5 open weights: what you can download and run
§ 01Repositories
GLM-5 models with public weights
Only models with an official Hugging Face repository are listed. Parameter counts appear only where the model card states them.
| Model | Hugging Face repository | Licence | Parameters | Inputs | Released |
|---|---|---|---|---|---|
| GLM-5 | zai-org/GLM-5 | MIT | 744B / 40B active | Text | Feb 11, 2026 |
| GLM-5.1 | zai-org/GLM-5.1 | MIT | Not published | Text | Apr 7, 2026 |
| GLM-5.2 | zai-org/GLM-5.2 | MIT | Not published | Text | Jun 16, 2026 |
| GLM-5.3 | zai-org/GLM-5.3 | GLM-5.3 (custom) | Not published | Text | Aug 18, 2026 |
| GLM-5.3 Flash | zai-org/GLM-5.3-Flash | MIT | 320B / 18B active | Text, Image, Video | Aug 26, 2026 |
Licences as shown on each Hugging Face repository. Always read the full licence file before commercial use.
API only, without public weights: GLM-5-Turbo, GLM-5V-Turbo, GLM-5.3 FlashX, GLM-5.3 Prime.
§ 02Background
What open weights mean for GLM-5
Weights, not a toy version
An open-weights release means the trained parameters themselves are public: you can download the files, run them on hardware you control, inspect them and fine-tune them. For GLM-5 these are the same models that Z.ai serves through its API, not cut-down demos. GLM-5's repository is accompanied by a technical report, "GLM-5: from Vibe Coding to Agentic Engineering" (arXiv 2602.15763), and a GitHub repository with deployment examples.
Open weights are not the same as fully open source in the strict sense: the training data and the complete training pipeline are not published. What you get is the model, its configuration, tokenizer and chat template, plus Z.ai's documentation on how to serve it.
MIT versus the GLM-5.3 licence
GLM-5, GLM-5.1, GLM-5.2 and GLM-5.3 Flash are released under the MIT licence. MIT is one of the most permissive licences there is: it allows commercial use, modification and redistribution, as long as the copyright and licence notice are kept.
GLM-5.3 is different. Its Hugging Face repository lists the licence as "other" with the name glm-5.3, which means a custom licence written by Z.ai. Do not assume it grants the same freedoms as MIT; read its terms, especially if you plan to offer a commercial service or ship derived models.
- Permissive (MIT): GLM-5, GLM-5.1, GLM-5.2, GLM-5.3 Flash.
- Custom licence: GLM-5.3.
- Not downloadable: GLM-5 Turbo, GLM-5V-Turbo, GLM-5.3 FlashX, GLM-5.3 Prime.
Which open model to start with
If you want the most capable open coding model of the family and its licence fits your use, that is GLM-5.3. If you need a permissive licence and image or video input, GLM-5.3 Flash is the natural choice: Z.ai reports that it outperforms GLM-5.2 while using only 18 billion active parameters out of 320 billion. GLM-5.2 is the MIT option when you want the flagship lineage with a solid 1M-token context, and GLM-5 is the best documented starting point for research.
§ 03Running locally
How local serving works
A general outline. The model cards and the frameworks' own documentation are the authoritative guides.
Pick a serving framework
GLM-5's model card lists vLLM, SGLang, KTransformers, xLLM and Hugging Face Transformers for local deployment. vLLM and SGLang are the usual choice for an OpenAI-compatible server on GPUs; KTransformers focuses on running very large mixture-of-experts models with CPU and GPU together; xLLM is the documented route for Ascend NPUs. The GLM-5.2 and GLM-5.3 Flash cards also link to Unsloth guides for local runs.
Serve, then call it like the API
Once a server is running, it exposes the same chat completions endpoint as the hosted API, so the code samples on our API page work against localhost. Use the tool-call and reasoning parsers named on the model card so that tool calls and thinking are separated correctly from the final answer.
Plan the hardware honestly
These are frontier-scale models. Even with mixture-of-experts sparsity, all weights must be stored in memory, and long contexts add a large key-value cache on top. The official GLM-5 examples shard the model across eight GPUs. Smaller setups rely on quantised community builds, which trade some quality for size.
Example from the GLM-5 model card
vllm serve zai-org/GLM-5 \
--tensor-parallel-size 8 \
--gpu-memory-utilization 0.85 \
--speculative-config.method mtp \
--speculative-config.num_speculative_tokens 3 \
--tool-call-parser glm47 \
--reasoning-parser glm45 \
--enable-auto-tool-choice \
--served-model-name glm-5These commands are quoted from the zai-org/GLM-5 model card. Flags and versions change; check the card and the framework release notes before you run them.
Frameworks named on the model cards
Hardware caution: we deliberately do not quote GPU or memory requirements. They depend on precision, quantisation, framework and context length — a 1M-token context (1.05M) needs far more memory than a short one. Size your hardware from the framework documentation and test with your real workload.
§ 04FAQ
Open-weights questions
Q01Which GLM-5 models can I download?
GLM-5, GLM-5.1, GLM-5.2, GLM-5.3 and GLM-5.3 Flash are published on Hugging Face under the zai-org organisation.
Q02Can I use GLM-5 weights commercially?
The MIT-licensed models (GLM-5, GLM-5.1, GLM-5.2 and GLM-5.3 Flash) allow commercial use with the licence notice kept. GLM-5.3 uses a custom licence, so check its terms first.
Q03Can I run GLM-5 on a laptop?
Not at full size. Even GLM-5.3 Flash has 320 billion parameters in total. Heavily quantised community builds can bring the memory need down, but the full-quality models are meant for multi-GPU servers.
Q04Are GLM-5 Turbo or GLM-5.3 Prime open weights?
No. GLM-5 Turbo, GLM-5V-Turbo, GLM-5.3 FlashX and GLM-5.3 Prime are available only through APIs.
Q05Is there an easier way to try these models?
Yes. Our free chat runs GLM-5.3 Flash and GLM-5.3 in the browser with no setup, so you can judge the models before investing in hardware.
Try it now
Ask GLM-5.3 something real.
No account, no card, no install. Type a question and watch GLM-5.3 Flash answer in seconds — switch to GLM-5.3 for the hardest problems.