October 24, 2026

Major features

Download a selection as one zip

Selection bars in Assets now have a Download button. Select several assets in the list, several files in a file view such as Outputs or Models, or files in a collection, and download them in one go. Delivery works like a single asset’s download: up to 2 GiB is streamed into one zip in your browser, and larger selections come as a list of per-file links. Viewers can select and download too.

Files that can’t be downloaded right now, such as linked sources, are left out, and the console tells you how many were skipped.

Read more →

New and refreshed official templates

• New templates: RustFS, S3-compatible object storage that takes MinIO’s place; TensorRT-LLM, NVIDIA’s inference server behind an access key; Fizgig, a LoRA training studio; Python 3.15; and Ubuntu 26.04.
• NVIDIA Triton Inference Server 26.09. The inference API now needs the generated API key shown under Connect, and the metrics port stays internal. /models is treated as your working folder, so connect storage there to keep your model repository.
• JAX 26.09, with support for the latest NVIDIA GPUs.
• Databases on pinned long-term versions instead of latest: PostgreSQL 18, MySQL 9.7, MariaDB 12.3, MongoDB 9.0, Redis 8, RabbitMQ 4 and QuestDB 10.1. A restart never moves a new Pod to a different major version.
• Updated versions of Elasticsearch, Label Studio, Text Embeddings Inference, Kokoro TTS, Ultralytics YOLO, Scientific Python, Polyglot Data Science, Pangeo PyTorch, RAPIDS Notebooks and ROS 2 Jazzy.

The versions they replace leave the template catalog. Pods already running them are not affected. MinIO is retired because its images are no longer published.

Improvements

• MOSS-Image uses the latest official workflows. Text to Image has a rewritten prompt enhancer, and Layout to Image now opens with example boxes. The set uses a smaller text encoder and VAE and is now about 13 GB. MOSS-Image Pods you already run download the new files (about 7.6 GB) once.
• More templates show their official logo on the template cards.
• Custom templates are easier to get right. The Connect button is now picked only from external ports. Image environment variables with an empty default are no longer forced to be filled in. The endpoint option is a single HTTPS toggle. The SDK refuses template ports that would put Connect on an internal port.

Fixes

• Security on by default for new Elasticsearch and Redis Pods. Elasticsearch now requires the generated password, and Redis now enforces the password shown under Connect. Pods created from the earlier templates keep their old settings. If you run one, we recommend redeploying from the updated template.
• Quick Deploy works again for recently renamed model repositories. Sets whose Hugging Face repositories were renamed, including Z-Image Turbo, Qwen-Image, Wan and FLUX, no longer fail with “Hugging Face did not respond”.
• Pods with internal-only ports no longer sit at Preparing. Internal ports have no address to check from outside, so the Connect tab now shows them with a neutral marker instead of a waiting state. QuestDB Pods now reach Ready.
• Writes during shutdown are kept on more kinds of Pod changes. The October 18 fix for deleting a Pod now also covers editing a Pod to detach a volume or move it to another machine, a spot Pod being reallocated, and a move that fails or is cancelled. Apps that save a final checkpoint after the shutdown signal no longer lose it on network or encrypted volumes.
• Reclaimed spot Pods wait when no machine is free. If a spot Pod is reclaimed and no other machine has room for it, it now stops and waits, then starts again on its own when capacity frees up. Before, it kept running. Demand vs. spot →