October 3, 2026

Major features

Watched folders — you choose what a Pod saves

A Pod saves its outputs by watching folders for new files. You now decide which folders are watched, per Pod.

• At deploy time, the Pod create form has a Watched folders section. Turn a template’s folders off or on, edit their include patterns, or add folders of your own.
• On a running Pod, open its Assets tab → Watched folders and change them live. No restart is needed; a stopped Pod picks the change up when it next starts.
• Folders on network storage are not watched. You manage that storage yourself and several Pods can share it, so saving from it would upload the same files more than once. Such a folder shows a Network storage badge and can’t be turned on, and a folder of your own on network storage is rejected at deploy. Folders on local disks beneath it are still watched.

Pods created before this release keep their template’s default folders until their next restart.

Watched folders →

A redesigned Pod creation wizard

Creating a Pod is now three clear steps with a progress bar you can click back through.

• Template: one screen with popular templates, and an All templates tile that opens the full Official / Custom list with categories and search.
• Configure: the versions of a template appear in a single list (version, CUDA, where it runs, image). Required environment values are filled in under Advanced settings, and Continue stays disabled until they are.
• Hardware: GPUs are grouped into Available, Popular, Data center, Workstation and Consumer tabs, with CPU cards, Pod type and pricing choices that explain why an option is unavailable, and the lowest GPU price in the bottom bar.
• Assets: inputs and outputs sit side by side. Input shows what you attached as chips (LoRA, Checkpoint, Images, ControlNet, Upscale Model); Output shows Save to and Save from (your watched folders) with presets for all files, images or videos.

Quickstart →

Serving runs on vLLM 0.30

New replicas of your serving deployments start on vLLM 0.30. Models load more reliably as a result:

• GGUF models load even when the repository has no separate config or tokenizer files.
• Embedding models run in pooling mode.
• DeepSeek-OCR family models start correctly.

About 40 older architectures that the new engine no longer supports can’t be newly registered as custom models. Replicas that are already running keep running; a new replica of an existing registration starts on the new engine.

Models → · Serving →

API changes

• Image, video and embedding requests are handled by your own deployment. Use your deployment’s endpoint URL. These requests on the shared base host (inference.meshive.ai/v1) now return 404. On your own URL model is optional; if you send one that doesn’t match the deployment, the request fails with 400 and code model_mismatch.
• A stopped deployment answers 503 immediately for synchronous and asynchronous requests. Previously a replica that hadn’t been removed yet could still answer, and asynchronous requests were still accepted.
• Asynchronous image requests are always queued. They are accepted with 202 even when no replica is ready or the deployment is busy; synchronous requests still return 429 when saturated. Repeating an asynchronous request with the same idempotency key returns the original 202, even after the deployment has been stopped or deleted.
• Webhook URLs must point to public addresses. Addresses on private or internal networks, including their IPv6 forms, are rejected when you save an endpoint. Webhooks →
• Loading an image from a private-network registry is rejected with 400. Other registry failures — an unexpected response, a certificate problem, a rate limit — now come back as specific 400 messages instead of a generic “Check your network and try again”.
• Mounting storage at /usr/local or below is not allowed.
• The legacy single-Pod creation endpoint (POST /v1/kubernetes/pod) has been removed. Create Pods through the console or the SDK.

API reference →

Improvements

• Deploy cards show how input downloads are going. Progress appears for every kind of storage, including on start and restart. A download that is verifying files or waiting for storage space says so instead of looking stalled, a failed download shows a collapsible excerpt of its log, and cards keep a stable order while they refresh.
• Sizes use KiB, MiB and GiB. The console, emails, SDK and error messages now label 1024-based sizes the way large clouds do, so sizes are no longer labelled GB when they are really GiB. GPU VRAM stays in GB, and network speeds are in decimal Mbps.
• Premium options match their descriptions. The internet premium requires a public connection averaging about 850 Mbps; the earlier check let almost any measured machine through. Super-Safe Host uses the Platinum uptime requirement shown in the console, and Latest CPU means the newest generation released for that CPU family. Surcharges now show their real percentage. Uptime and fees →
• Placement is a little smarter. For GPU Pods, ties between otherwise equal machines now go to newer memory and CPU generations and to faster disks, and older CPUs are no longer ranked ahead of newer ones. Node matching →
• A Pod whose original machine can no longer run its configuration starts on another machine by default, with the reason shown, instead of waiting in a loop for a machine that can’t take it.
• Faster name lookups inside Pods. Names with two or more dots resolve with far fewer retries. On Alpine-style (musl) images, a partially qualified cluster name such as service.namespace.svc may no longer resolve; use the full service.namespace.svc.cluster.local.
• Serving deploy previews match real placement. The GPU count and price shown at deploy are computed from each machine’s actual VRAM.
• New ComfyUI Pods are password protected. Jupyter and code-server now require a password that is generated when the Pod is created and shown in Connect. New ComfyUI Pods use CUDA 13.0 only; existing Pods keep running as they are.
• The Blender template is pinned to 5.2.2. The unpinned latest image failed in headless rendering and rendered on the CPU by default.
• SDK, CLI and MCP. Writes recover safely from retries, moving a Pod asks for data-loss consent separately, serving deploys confirm cost, task logs page properly, and meshive transactions (also an MCP tool) shows why a Pod is still creating. SDK and CLI →
• A more consistent console. Notifications share one look, and a lost connection now reads “Connection lost” instead of a browser message. Deploy and Create buttons are larger, and the first-load screen is animated.
• A refreshed meshive.ai. The landing site has a redesigned hero and new pages by solution and industry, and its storage and memory figures use GiB and TiB like the console.

For hosts

• Restarts keep your machine listed. Restarting the agent or rebooting no longer drops a listed machine to unlisted. A machine stopped by a GPU error stays unlisted when it reconnects, so it isn’t offered again until it has recovered.
• You are told when a machine is unlisted automatically. An email now follows an automatic unlist for boot errors or warnings, a failed port check, GPU errors or a PCIe mismatch, with how to relist. The email sent when maintenance ends now says the machine returns unlisted. Machine states →
• Maintenance windows respect machine state. A machine that is being deleted or onboarding can’t register or start a window, and windows that never started are cleared when a machine is removed. Maintenance →
• Disk speed is measured where Pods store data. The test runs on a temporary volume inside the platform’s storage group instead of the root partition, repeats weekly, and a machine that was never measured shows — instead of 0.
• Hardware is detected more accurately. Data-center GPUs such as HGX systems are counted as GPUs, so they are no longer registered as CPU-only hosts; NVLink is counted by active links; GPU model and VRAM match the price table; and older CPU models are no longer read as the newest generation.
• Re-verification after a boot problem is safer. It compares CPU threads, checks compatibility when a machine returns, and no longer stops Pods that are being moved or waiting for another machine.
• Network storage nodes recover on their own. NFS serving is checked and repaired automatically, and deleting a volume no longer gets stuck on leftover client state.
• A skipped runtime configuration step no longer unlists machines. It was being reported as a boot error, which dropped listed machines to unlisted at every agent start. Machines also remount their volumes at boot when the database is unreachable.
• Two changes from September 18 are now live for everyone. Reference GPU prices refresh every week, and the output of a failed setup step is recorded in the machine history.

Policy changes

• Workspace roles are enforced more strictly. Effective immediately:
  • Viewers can no longer create, edit or delete custom templates, or rename Pods, storage, templates, registries or workspaces.
  • Registry logins and storage-bucket credentials can be added or removed only by Admin and Billing members.
  • Deleting an inference result or cancelling a request is limited to the person who made the request, Admin and Billing.
  • Billing details of a workspace are visible to its members only.
  • Official templates can be changed only by Meshive administrators.
• Storage is billed per GiB. Managed asset storage, model cache and linked external model storage now count in GiB (2³⁰ bytes) at the same unit price, so the same data is billed about 7% lower. Asset billing →
• GPU Pod extra vCPU and RAM have one flat rate. Anything above the amount included with the GPU is billed at a single rate that is the same on every machine, so the price shown at deploy always matches the charge. CPU Pods are placed on, and billed at, the CPU option you picked. Pricing →
• CPU Tasks run on the cheapest CPU option shown in the submit form and are billed at that option’s rate. If the price moves above the limit you set, the submit fails with 409 and the new price. Tasks →

Fixes

• Declining the data-loss warning is respected. If you decline to start a Pod on another machine because its workspace data would be lost, the Pod now stops. Previously it could still be moved later without your consent, and the data was removed.
• A failed machine move rolls back cleanly. The Pod left behind on the new machine is shut down instead of lingering, a Pod you deleted mid-move isn’t brought back, and the next attempt waits about ten minutes.
• A failed Pod edit rolls back completely instead of ending in a false “rollback incomplete”.
• Starting a Pod on the same machine after it moved now reserves the machine it is on, not the one it left.
• Serving metrics and autoscaling count only your own workspace. Live metrics, queue depth and scale-up decisions no longer include other workspaces running the same official model.
• Failed starts are easier to diagnose. The cards no longer show a container that is still initializing as an error, and a final download report that finished is no longer lost, which used to leave an earlier failure message on screen.
• The log window no longer shows raw JSON fragments while a Pod is waiting.
• The pending-invite list no longer refreshes endlessly.
• Several smaller console fixes: the Pod drawer no longer opens the name in edit mode when you close it, the Recent outputs strip shows the newest outputs, and an empty Assets tab names the folders actually being watched.