Blog / Model guides
GLM-5.3 open weights: expected release date and what you will need to run it
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.
What you will likely need to run it
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. Treat this table as expectations, not measurements; we will replace it 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 |
| Context | 256K tokens |
| License | GLM-5.2 shipped under MIT; we expect the same and will confirm at release |
| Hardware class | A multi-GPU H200-class or B200-class shape. We will publish the exact shapes we qualify |
What we will measure on day one
When the weights drop, we pin the release, review the license, and run our qualification process. This page then gets the same treatment as every model guide we publish:
- Cold boot time, from fresh instance to ready endpoint, measured on a real deployment in our own cloud account
- Warm boot time from a prepared stage, where supported
- Throughput and time to first token, always published together with the workload they were measured under (4,096 input tokens, 1,024 output tokens, concurrency 1)
- Observed price per hour at each provider that can run it, and whether capacity was actually available when we checked
How to be ready
If you want to run GLM-5.3 privately the week it lands, the only slow step you can do in advance is connecting your cloud account. Once the model reaches the catalog, deploying it is a few clicks: pick the shape, set a budget cap, confirm. Your prompts stay on your own instance.