October 10, 2026

Major features

More RAM included with every GPU Pod — same price

GPU Pods now include 1.25 GiB of RAM per GB of VRAM (up from 1 GiB per GB), rounded up, at the same GPU rate. The maximum stays at 2 GiB per GB of VRAM.

• New Pods get the larger amount automatically — an RTX 5090 Pod now starts with 40 GiB instead of 32 GiB.
• Existing Pods keep their current RAM and price. When you next edit one, RAM you added on top is measured against the larger included amount, so the price never goes up.

Pricing & Instances →

ComfyUI RAM saver mode and one-click fixes for out-of-memory

• ComfyUI Pods have a RAM saver mode switch. It runs ComfyUI without keeping intermediate results in memory between runs, so heavy workflows that used to run out of memory fit in the included RAM, at the cost of recomputing some steps on the next run.
• When a Pod is stopped because it ran out of memory, its card now offers Turn on RAM saver mode (ComfyUI) and Increase RAM. Each opens the edit form at the right place.

Storage Connect explains itself

Connect Storage opens an access path to a volume without running a Pod.

• The dialog explains what to use it for and its limits.
• SSH sessions start in your volume’s folder, and tool caches are kept inside the volume.
• Large downloads (for example hf download) get more memory headroom.
• If the access path restarts — for example after running out of memory or filling its small temporary disk — the volume card shows a Restarted badge with the reason and how to avoid it.

Storage →

Improvements

• Edit attached storage in place. In a Pod’s Storage table, each attached volume has an edit button. Change its mount path, or for a new local disk, its name and size too. The X button now reads “Detach — the storage isn’t deleted”. A changed mount path also stays put: a later edit, such as adding RAM, no longer moves the volume back to its old path.
• Deleting outputs is final. Deleting every file in an output removes the whole item instead of leaving an empty “0 files” row. Files you deleted no longer come back after the Pod restarts, is edited or stops.
• Saving outputs is more robust. A Pod that saves its outputs now skips only the individual files the platform can’t accept — including files whose names aren’t valid UTF-8 — instead of holding back the whole folder.
• Assets downloads keep their file names, and Open in new tab opens files in the browser instead of downloading them.
• Long non-streaming requests to your serving deployments can now run for up to about 10 minutes instead of failing after about 2. A request that takes longer returns a clear 504 that suggests stream: true or a lower max_tokens.
• Clearer messages when credit runs out.
  • Deploying serving with no credit shows Credit required with an Add credit link.
  • Messages about starting Pods, storage and SSH say whose billing account is out of credit and what to do.
  • The Stop dialog notes that storage keeps being billed after a Pod stops.
• Inviting someone who is already a member of the workspace shows “Already a Member” instead of a generic error.
• SDK, CLI and MCP.
  • A Pod that failed to create now reports why, instead of looking deleted. Lifecycle calls on it return pod_creation_failed.
  • Pod connection details, downloading assets and outputs, registering serving models, importing assets, watched folders and SSH are available.
  • SDK & CLI →

For hosts

Fan curve

Each GPU machine has a new Fan curve tab.

• Default curve. The agent drives the GPU fans on a platform curve that reaches full speed well below throttling temperatures. Fans now run at about 30% or more even when idle, so machines run cooler and somewhat louder.
• Your own curve. Drag points to set a custom curve. It must reach 100% at 90°C or below.
• NVIDIA driver (auto). Choose this to leave the fans to the driver — including stopping them at idle — or if you run your own fan tool.
• The tab shows live temperature and fan speed for each GPU.

A new internet speed test

The internet speed test no longer uses a third-party speed test service.

• How it measures. It downloads a large public image layer from Docker Hub and uploads test data to Meshive storage.
• When it runs. At start-up and at a random time within the weekly window. It also re-measures when the network card renegotiates a different link speed.
• Fairness. It waits while the line is busy, and keeps the better of up to two rounds.
• Machine details now show the link speed and when the line was last measured.

Security → · Node matching →

Network diagnostics

Diagnostics → Network lists each network card’s link speed and the fastest speed it supports. It also warns about problems such as a link running below its maximum, half duplex, or a link that keeps dropping.

• Agent updates now wait until a machine has finished starting up. An update can no longer interrupt the start-up checks, which could leave a GPU listed as unsupported.

Policy changes

• One credit balance. There is no longer a separate free-credit balance.
  • There is no sign-up credit — it has been $0 since July.
  • Referral rewards are added as regular credit. It works for GPU Pods, tasks and Serverless and doesn’t expire.
  • Credit earned by sharing a serving deployment’s idle capacity is also regular credit.
  • Balances you already have are unchanged.
  • FAQ →
• Privacy Policy updated (October 10, 2026) to describe workload console logs — the output your Pods, tasks and endpoints write to standard output and standard error. It also describes their retention for about 30 days. Privacy Policy →

Security

• Sign in again once. Sign-in sessions now use a stricter cookie that is sent only to the Meshive API. Everyone needs to sign in once after this release; after that you stay signed in as before.

Fixes

• Editing or restarting a Pod no longer sends offline and online emails. The Pod also no longer briefly shows as running and then back to starting.
• Fairer billing around edits. Time spent pulling a new image during an edit or restart is no longer billed. An edit that fails and rolls back no longer bills the time the Pod was down.
• Daily host payouts are processed more robustly.