v1.26 release notes, 21 September 2026
What changed in release v1.26 of zCLOUD, the GPU cloud: who each change applies to and whether a step is needed.
**Release window:** 21 September 2026, 14:00 to 18:00 UTC.
**Service impact announced for the window:** sign-in possibly unavailable for a period; existing GPU tasks not interrupted.
How to read an entry
Each entry names the product and who it applies to.
| Label | Meaning |
|---|
| Action required | A step is needed to receive the change or to keep working |
| Changed | Existing behaviour differs |
| Added | A new control or capability |
Action required
**Action required. Changed** (zCLOUD). Applies to: users who create tasks. Newly created training, fine-tuning, inference and Web IDE tasks mount only the storage chosen in the **Storage** section of the create form. If no storage is chosen, none is mounted. Tasks created before the update are not affected. A workflow that relied on storage being mounted without choosing it needs the storage chosen when the task is created.
Added
- **Added** (zCLOUD, inference). Applies to: developers who create inference tasks. A multi-role mode, also called PD-disaggregated serving. One inference task can run one to five roles, such as prefill, decode and router. Each role has its own image, start command, resource specification and replica count. The release note states the aim as higher throughput for long-context and high-load workloads. The **Inference Mode** toggle offers **Standard Mode** and **Multi-Role**.
- **Added** (zCLOUD, inference). Applies to: developers who create inference tasks. Session affinity: requests in the same conversation are handled preferentially by the same replica. It applies to **Standard Mode** inference tasks with more than one replica. The configured value must be lower than the number of replicas. No client change is needed.
- **Added** (zCLOUD, inference). Applies to: developers who create inference tasks. A custom **Model Name** can be set when an inference task is created. It affects the `model` parameter passed in API calls, and it is carried over when a task is cloned.
Changed
- **Changed** (zCLOUD). Applies to: all users. Repeated error messages in the console are merged to reduce stacked messages.
Next steps
- [Storage and PVCs](https://docs.zettabyte.com/docs/zcloud/models-data-storage/storage): how storage mounts into a task, and the storage limits.
- [Deploy a model](https://docs.zettabyte.com/docs/zcloud/deploy-and-serve/deploy-a-model): create an inference endpoint and read the Model Name column in the Inference list.
- [Release notes](https://docs.zettabyte.com/docs/release-notes): all releases, newest first.