August 16, 2026

This release adds GPU memory selection to Pod deploy, raises the per-file upload ceiling, and lands a broad set of fixes across the Pod lifecycle — editing, restarting, logs, and how download failures are reported.

Major features

Choose GPU memory, not just the model

Cards of the same GPU model don’t all carry the same memory. Pod deploy now shows GPU memory (VRAM) alongside the model — filter by capacity and pick the exact card you need — and your Pod is placed on a card with exactly that memory, priced for that memory. When you edit a stopped Pod to a capacity no card currently provides, the message now says exactly that, instead of wrongly reporting the GPU model itself as unavailable. Read more →

Improvements

• Uploads up to 500 GiB per file. The per-file limit for storage uploads was raised from 100 GiB. Upload sessions also no longer expire while data is still flowing — only idle time counts against the session — so a slow connection can finish a very large upload without racing a deadline. Read more →
• Input download failures explain themselves. When a Pod’s input file can’t be fetched — an expired download link, for example — the Pod now shows which file failed and why. The notice stays visible until you dismiss it (and dismissing now works reliably), stale failure cards no longer linger after an edit, and long file names are shortened in the middle so both ends stay readable.
• No-change edits do nothing. Pod Edit’s Update button is disabled until something actually changes, and submitting an edit with no changes no longer restarts the Pod.
• Output files reach your storage sooner. Files a running Pod writes to its output folder are now picked up into your storage much sooner after they are written.

Fixes

• After a Pod restarts, its log view starts fresh — lines from the previous run no longer stay mixed into the new session’s logs.
• The live log panel could stop updating once its scrollback filled, show duplicated lines, or go quiet while still being watched — all fixed.
• Changing the GPU model of a stopped Pod now takes effect reliably instead of being silently lost, and the shown price follows the newly selected GPU.
• The indicator for whether a stopped Pod can start again is more accurate: it checks capacity across all eligible machines instead of a narrower set, and the warning shown when the whole cluster is full no longer wrongly blames the Pod’s original machine.
• Fixed a permissions gap that could make Pod storage volume operations fail with an authorization error in some workspaces.
• Editing a Pod no longer regenerates its auto-created credentials — passwords and tokens issued at deploy keep working across edits.
• Renaming an endpoint no longer causes an unnecessary container restart, and endpoint lists keep a consistent order.
• A Pod edit that failed partway could leave the Pod stuck on a storage mount error; cleanup now runs in the correct order so the Pod recovers.
• Edits that moved a Pod to a different machine could fail right after its input files finished downloading; both causes are fixed.
• Editing a Pod created from a template whose input requirements had since changed is no longer rejected.
• The folder where a Pod’s saved output files are collected no longer moves on the Pod’s first edit — outputs keep landing in the same place for the Pod’s whole lifetime.