October 12, 2026

Major features

Environment variables: your MESHIVE_ names work, platform variables move to MESHIVE_SYS_

• Variables you add are kept. A variable whose name starts with MESHIVE_ — for example MESHIVE_COMFY_TOKEN — now reaches your Pod’s container. Before, such variables were silently removed. If an earlier Pod needed one, add it again in Edit.
• Platform variables have their own prefix. The variables Meshive sets for you now start with MESHIVE_SYS_, such as MESHIVE_SYS_OUTPUT_DIR and MESHIVE_SYS_DATASET_DIR. The older names still hold the same values for now, so existing scripts keep working — switch to the MESHIVE_SYS_ names when you can. An existing Pod gets the new names the next time you edit it.
• Only Meshive’s own names are reserved. MESHIVE_SYS_* and the older platform names can’t be used for your own variables. The form marks such a row with “Meshive sets this variable — rename this key”, and the API returns 400.
• Read-only, out of the way. In a Pod’s edit form, the platform variables sit in a collapsed, read-only group inside Environment variables, so the list you edit holds only your own variables.
• Tasks follow the same rule. Your script can read its task ID from MESHIVE_SYS_TASK_ID (the older MESHIVE_TASK_ID holds the same value).

Importing & creating assets → · Tasks →

Tasks with model inputs can use more GPUs

• Tasks whose inputs are Hugging Face, CivitAI or URL assets — including gated and private ones that use your saved token — are no longer limited to a smaller set of GPUs.
• The GPU list and price quotes now show only hardware a task can actually be placed on right now, so a quote no longer succeeds only for the submission to be rejected.
• When a submission can’t be placed, the message says why and what to do — the hardware is busy for a few minutes, isn’t available for tasks at the moment, or is available only for public container images. The error also includes a code you can handle in scripts.
• If an input can’t be downloaded, the task now ends right away with the reason instead of waiting for a start-up timeout.

Tasks →

Linked sources are checked before you deploy

• When you deploy a Pod, submit a task, or add a linked asset to a Pod in Edit, Meshive first checks that each Hugging Face, CivitAI or URL source is still there. If it was deleted, made private, or your saved token can’t access it, you get a clear Linked Asset Unavailable error up front instead of a failed start. Temporary network problems don’t block the deploy.
• Temporary links are refused. A signed download link that expires can’t be registered as a URL asset, because linked files are downloaded again every time a Pod or task starts. Use the model’s page URL (Hugging Face, CivitAI) or a permanent link instead.
• Correct sizes for compressed downloads. URL assets from servers that compress downloads now record the real file size. If a Pod or task reports that a text-type file you registered earlier is larger than expected, register it again.

Import from Hugging Face or CivitAI →

Improvements

• Starting or stopping a Pod that is already changing state now returns a clear conflict message instead of “Pod not found”.

For hosts

A faster, more accurate internet speed test

• The test now measures against the nearest Meshive speed test endpoint first. When that reaches about 85% of the network card’s link speed, the other sources are skipped, so most tests finish in seconds.
• Otherwise it also measures Docker Hub and Meshive storage and keeps the best result. The Docker Hub download is warmed up first, so a cold cache no longer understates a fast line.
• Multi-gigabit machines far from Korea will see a higher, more accurate upload speed.
• Your first measurement with the new method happens at the next weekly test, a restart, or a link speed change.

Node matching → · Security →

Network self-repair

• Replaced or renamed network cards. On machines whose network is configured entirely by DHCP, if the configured network port disappears — for example after a card is replaced or renamed — the agent adds a DHCP configuration for the port that has a link. The machine comes back online on its own instead of staying offline. The agent writes only its own configuration file and never edits yours.
• Full link speed at boot. The agent asks each network card to advertise every speed it supports, unless you’ve limited it yourself, and reverts if the link doesn’t come up.
• No boot delay from unplugged ports. A configured port that has had no cable for a day no longer stalls boot.
• Machines with static IP addresses, bonds, VLANs or NetworkManager get diagnostics only.
• The agent’s firewall is restored automatically after a reboot.

Fixes

• A machine that is offline with only storage on it — no Pods running — no longer accumulates downtime. Downtime is still counted while a renter’s Pod is on an offline machine.