July 10, 2026

Our biggest release yet — four major features and a smarter serving engine.

Major features

Shared serving

Call models other teams host — no GPU to rent, nothing to deploy, billed per token with your existing API key. Browse what’s live under Serverless › Shared.

If you host a deployment yourself (official text models on vLLM), you can now flip Share idle capacity: guests use only your spare throughput, your own requests always take priority, and 100% of what guests pay is credited to your balance — a discount on the deployment you’re already running. Read more →

Encrypted storage

Any volume — local or network — can now be encrypted at rest, free, opt-in at creation. The key is never stored on the host machine, so the volume’s contents are unreadable outside their normal use. Encrypted network volumes unlock on demand through Connect and auto-lock when idle. Read more →

Tasks

Run a Python script once, to completion, on a matched GPU or CPU node — batch jobs, fine-tuning, preprocessing — with per-key secret environment variables, output downloads, and webhook notifications. Billing is by runtime, with automatic refunds for platform-side failures. Read more →

New image & video generation engine (ComfyUI)

Image and video serving now runs on a ComfyUI-based engine. With it:

• Video models are back in the official catalog — text-to-video and image-to-video, deployable in one click.
• GET /v1/models now reports supported_parameters per generation task, so you can see exactly which request parameters a deployment applies.
• Known limitation: registering custom image/video models is temporarily paused during the migration — existing custom deployments keep serving, official models are unaffected, and custom diffusion registration is on our roadmap to return. Details →

Improvements

• Smarter autoscaling & routing — LLM scaling decisions and request routing now track each replica’s real memory headroom more accurately: less overload on busy replicas, safer scale-down, and sticky conversations that heal themselves when a replica goes away.
• Billing transparency — the console shows exactly when billing is active: serving replicas bill from the start of the weight download, with image pull and warmup free; pods bill from the moment the container is running, and a pod can be stopped even while it’s still being prepared.
• Email alerts for hosts — you get an email when your machine can no longer start client workloads or a system component fails (machine states guide).

Fixes

• API errors are now consistently OpenAI-shaped, with clearer causes for upstream failures.
• Long video generations tolerate brief infrastructure interruptions instead of failing early.
• Logs for finished pods and tasks remain viewable after termination.
• Fixed cases where the official template list could fail to load.