Blog / Model guides
GLM-5.3 hardware requirements and expected open-weights release date
Updated 27 August 2026. The zai-org/GLM-5.3 page on Hugging Face now lists 28 August as the release date. The smaller GLM-5.3-Flash, which Z.ai had been running anonymously as Ox Alpha, was released under MIT on 26 August; see its own page. The hardware section below is calculated from the GLM-5.2 base; we have not deployed the model.
What GLM-5.3 is
GLM-5.3 is Z.ai's latest flagship, launched on August 14, 2026 on their API and coding plan. Z.ai says it keeps the same base model as GLM-5.2, with the capability gains coming from scaled-up post-training. Their headline claims are in coding, where they report it as the strongest open-weights system they have measured, and in long-horizon agentic work. The weights are not public yet: Z.ai is holding them for a safety evaluation, saying capability in cybersecurity tasks grew faster than they expected during training.
GLM-5.3 hardware requirements: what you will likely need
Because Z.ai says the base is unchanged from GLM-5.2, the published GLM-5.2 footprint is the best available guide to what self-hosting GLM-5.3 will take. Everything in this section is expected or derived, not measured; it will be replaced with measured numbers when the weights are out.
| Property | Expected (from the GLM-5.2 base) |
|---|---|
| Architecture | Mixture of experts, roughly 753B total parameters, roughly 40B active per token |
| Weights on disk, bf16 | Roughly 1.5 TB (derived: 2 bytes per parameter) |
| Weights on disk, fp8 | Roughly 750 GB (derived: 1 byte per parameter), if an fp8 checkpoint ships |
| Smallest single node, fp8 | 8x H200 (1,128 GB of GPU memory): the weights leave about 375 GB for KV cache and activations |
| Smallest shape, bf16 | Two nodes of 8x H200 (2,256 GB) or a Blackwell node such as 8x B300 (2,304 GB). A single 8x H200 node (1,128 GB) and 16x H100 80 GB (1,280 GB) are both too small for 1.5 TB of weights plus cache |
| Context | 256K tokens |
| License | GLM-5.2 and GLM-5.3-Flash shipped under MIT; Z.ai has not stated the flagship's license |
A single-node deployment depends on an fp8 or fp4 checkpoint. Z.ai has not said whether one ships alongside bf16. Per-token compute is set by active parameters, not total, so a 40B-active model is cheaper to serve per token than its total size suggests. The cost of self-hosting it is set by holding 750 GB or more of weights in GPU memory for every hour the deployment runs.
The Flash variant arrived first
The loudest request on the GLM-5.2 repository was a smaller Flash or Air variant that fits a single GPU or a workstation. On 26 August Z.ai released GLM-5.3-Flash, the model it had been running anonymously on OpenRouter as Ox Alpha, under MIT. It has 320B parameters with 18B active, and its native FP8 checkpoint is about 306 GiB, so it needs at least 4x H200 or 8x H100 80 GB: a smaller node than the flagship requires, but not a single GPU. The details are in GLM-5.3-Flash hardware requirements and license; the attribution history is in Is Ox Alpha open weights?
Running it privately
The one step that can be done before the release is connecting a cloud account. Once the model is in the catalog, a deployment is: pick the shape, set a budget cap, confirm. Prompts stay on the instance in your own account.
Frequently asked
Is GLM-5.3 open source right now?
Not yet. As of 27 August 2026 the zai-org/GLM-5.3 repository on Hugging Face is a placeholder that lists 28 August 2026 as the release date. Z.ai launched GLM-5.3 on its API on 14 August 2026. GLM-5.2's weights shipped under MIT, as did the smaller GLM-5.3-Flash on 26 August; Z.ai has not stated the flagship's license.
When are the GLM-5.3 weights expected?
28 August 2026, the date on the zai-org/GLM-5.3 page on Hugging Face. That is Z.ai's own listing, two weeks after the API launch, and it can still slip.
Is GLM-5.3-Flash the same thing as GLM-5.3?
No. GLM-5.3-Flash is the 320B-parameter, 18B-active model Z.ai previewed anonymously as Ox Alpha and released under MIT on 26 August 2026. GLM-5.3 is the flagship on the roughly 753B GLM-5.2 base. Flash needs at least 4x H200; the flagship needs an 8x H200 node at fp8.
Did Z.ai change the base model from GLM-5.2?
Z.ai says no: GLM-5.3 keeps the GLM-5.2 base, with the gains coming from scaled-up post-training. That is why the GLM-5.2 footprint is the best available guide to the hardware.
How many GPUs will GLM-5.3 need?
Derived from the GLM-5.2 base: roughly 753B parameters is about 750 GB at fp8, which fits a single 8x H200 node (1,128 GB) with room for KV cache, or about 1.5 TB at bf16, which needs two 8x H200 nodes or a Blackwell B300 node. The exact shapes depend on which checkpoint precisions Z.ai ships.
Will LLM Hangar support GLM-5.3?
GLM-5.3 is added to the catalog after the weights are released and the license reviewed. Cold boot, throughput and price per hour are measured on a deployment in our own cloud account before they are published here.