September 2, 2026

Major features

Templates that bring their own storage

Official templates no longer ask you to attach storage before deploying. Each working directory a template declares now gets its own volume, provisioned automatically when the Pod starts and growing as you fill it. If you mount a storage volume that already covers a path, that volume is used instead — nothing is provisioned twice. Auto-provisioned volumes are billed by allocated size for as long as the Pod lives, and the deploy summary lists every path that will get one. Storage →

Include rules: choose what gets harvested

Harvested paths can now carry include rules — comma-separated glob patterns like *.safetensors, images/. Only matching files are collected into Assets; an empty list keeps collecting everything, so nothing changes until you opt in. Set them per template under Semantic Paths, or override them per Pod from its asset settings. Harvesting →

The AI-Toolkit template puts this to work: training runs now harvest only the trained adapter (*.safetensors plus its config) and register it under the Adapters tab — ready to select in a ComfyUI LoRA slot. Assets harvested before this release stay where they are under Outputs.

Read-only API access with your Meshive API key (SDK 0.0.7)

A read-only API surface is now available for Meshive API keys with the READ scope: workspaces, members and storages, Pod and machine metrics, GPU availability, credit and earnings history, templates, servings, tasks, and your Assets — everything needed to script dashboards and automation without a browser session. Secret values never ride along: template and task secret env values are masked, billing history omits receipt links, and API keys are listed by prefix only. Use it from SDK 0.0.7 or plain HTTPS. SDK & CLI →

Task outputs in Assets, within seconds

Task results used to appear in Assets on a fixed polling cycle — up to a minute after the task finished. Delivery is now event-driven end to end, so outputs typically register within a few seconds of the task completing.

Improvements

• Placement preview — the Pod asset picker now shows the exact destination path for every file group before you deploy, so you can see where things will land instead of finding out inside the container.
• Denser task list — task cards are about half their previous height, so roughly twice as many tasks fit on screen.
• Input asset rows show an “n of m files” count, making partial file selections visible at a glance.
• The template form’s Semantic Paths section now explains itself in place — hover any column header for what it controls and its default.
• Catalog cleanup: outdated duplicate variants of several official templates (older ComfyUI preset builds and a legacy Unsloth image) no longer appear in the catalog. Pods already running on them keep working.

Policy changes

• Exclude patterns retired. Harvest filtering is now include-based. If you had exclude patterns configured on a template or Pod, they no longer apply — harvesting collects everything for those paths until you set include rules.
• Undeclared asset roles are rejected. Connecting an asset under a role the template doesn’t declare now fails with a clear message at deploy, instead of silently landing the files in the default input folder. Pick a declared role, or set a target directory explicitly.

Fixes

• The Unsloth template starts reliably — its studio UI now opens on the Connect button’s console port, with Jupyter available alongside.
• Outputs tab multi-select works again in the asset file pickers.
• The asset file search box no longer drops keystrokes while results load, and file search is now case-insensitive — searching PNG finds png.
• Long filenames in the file pickers are easier to read — wider columns, with the full name shown on hover.
• Files selected for a Pod now land in the same destination the picker showed — re-editing a selection no longer forgets its target directory.