Changelog

New features, improvements, and fixes across self-hosted CosmicAC, with the newest release first. This changelog follows Keep a Changelog, and CosmicAC follows Semantic Versioning.

1.3.0 - 2026-10-27

Adds team storage volumes. A storage volume is persistent storage that a team attaches to a GPU Container Job. The volume keeps its data after you delete the job. You can create, list, and delete volumes in the web interface or with the CLI. For more information, see Use team storage volumes. The web interface no longer offers Google sign-in.

Platform

Web

Added

• View team storage volumes on the Storage volumes settings page, where team leads and platform administrators can also create and delete them
• Attach a team storage volume when you create a GPU Container Job, so that the job's data in /mnt/team-volume remains after you delete the job
• Receive an email when a team lead or platform administrator suspends you from a team or changes your role in it

Changed

• On a new deployment, only members, team leads, and platform administrators can see a team's member list and details
• Notifications for a failed job include the error message that explains why the job failed
• Job metrics charts add each new data point to the existing chart instead of rebuilding the chart

Removed

• The option to sign in and sign up with a Google Account

Fixed

• The Specification section of a GPU Container Job shows the node that the job runs on
• If you're already signed in, a password reset link or a CLI sign-in approval link opens its own page instead of the home page
• The dashboard counts a new job's GPUs as used when you create the job, not only after the cluster places the job on a node

CLI

Added

• List team storage volumes, and create or delete them if you're a team lead or platform administrator
• Attach a team storage volume to a new GPU Container Job, at the interactive prompt or with a command-line option
• List only the teams that you can switch to, which excludes archived teams and teams where your membership isn't active

Fixed

• A streamed chat with a single message exits when the response ends, instead of waiting for you to press Control+C

API

Added

• Create, list, and delete team storage volumes
• Attach a team storage volume by its ID when you create a GPU Container Job
• Sign out of your current browser or device only, or of every browser, device, and CLI that uses your account
• Pass a setup script that runs once when the container of a GPU Container Job first starts, for example to install packages
• Pass a startup script that runs each time the container of a GPU Container Job starts, for example to start a service
• Get the details of one Managed Inference model, such as its status and health, without listing every model
• Check whether a Managed Inference endpoint requires an authorization header, from the model list or from a single model's details

Deployment

Added

• Turn on team storage volumes on a Gcore rack by setting the rack's Gcore project, region, and API token
• Run a read-only deployment task that checks whether a rack can serve storage volumes and whether the orchestrator has registered the rack
• Limit how many storage volumes each team can hold, including failed volumes, with a default of 5
• Set one token secret that app-node, the inference proxy, and the monitor all use to sign and verify team tokens
• Deactivate an account for the whole deployment, which ends the account's sessions and blocks its sign-in and password reset
• Reactivate a deactivated account so that the person can sign in again
• Set the session lifetime in the app-node configuration, with a default of 12 hours from sign-in

Changed

• Upgrading adds the storage volume permissions to the roles of an existing deployment, so team members can use storage volumes
• The monitor enforces team-scoped access by default, so users see logs and metrics only for the jobs of their own teams
• Service images run Node.js 24.22.1, which fixes a known Node.js vulnerability

1.2.0 - 2026-10-07

Adds teams and role-based access, so each team has its own resources. Adds sign-in to the web interface and the CLI, recommended model configurations that you manage in the web interface, and vLLM model registration from vLLM Recipes. CosmicAC hashes API keys with a stronger algorithm, so CosmicAC might reject API keys that you created before 1.2.0. To replace a rejected key, create an API key in the web interface.

Platform

Web

Added

• Sign in and sign out, and approve a CLI sign-in from the browser
• Work in teams, switch the active team, and accept an invitation to join a team
• Create, update, and archive teams, and invite, suspend, reactivate, and remove members
• Give each member a role, platform administrator, team lead, member, or viewer, that sets what they can do
• Reserve nodes for a team, and see the policy for unreserved nodes, on the Nodes settings page
• Review user activity on the Activities settings page, with filters
• Register a vLLM model, and prefill its configuration from vLLM Recipes
• Add, edit, filter, search, and archive recommended model configurations

Changed

• Downloading job logs exports the whole History window with the current filters
• Clicking a job row opens the job, and the jobs list sorts by column

CLI

Added

• Sign in with cosmicac login, sign out with cosmicac logout, and show the signed-in user with cosmicac whoami
• Manage teams and members with cosmicac teams, and switch the active team with cosmicac teams use
• Check a team's reserved capacity with cosmicac teams capacity
• List the jobs of every team with jobs list --all-teams, for platform administrators

Changed

• Job, API key, and model commands act on the active team
• inference chat and inference transcribe check the API key before the session starts
• jobs shell resizes the remote terminal with your window
• api-keys usage masks the key values

API

Changed

• CosmicAC hashes API keys with PBKDF2-HMAC-SHA512 instead of HMAC-SHA384, so CosmicAC might reject API keys that you created before 1.2.0

Fixed

• Model health history counts 4xx client errors as failures, to match the endpoint's success rate
• An endpoint no longer stays Degraded after probes fail while it warms up
• A chat completion rejects a video_url that points to a local or reserved network address

Deployment

Added

• Multi-architecture service images that also run on arm64 hosts
• Set the policy for unreserved nodes with WRKORK_COMMON_CONFIG__node_reservations__default_policy

1.1.0 - 2026-09-02

Adds multi-node Managed Inference to both clients, an observability aggregator that CosmicAC now runs as part of the deployment, and job logs and metrics history in the web interface. One change is breaking. --api-key no longer accepts a value.

Platform

Web

Added

• Serve a Managed Inference model across two nodes with a 16-GPU count
• Read live and historical job logs, with time windows, search, and download
• View job metrics history as charts on a job page
• Run a health check on a Managed Inference endpoint from the Model health page
• Set up job notifications, preview the payload, and review recent deliveries
• Connect CosmicAC to your Prometheus and Loki from an observability settings tab

Changed

• Location moved from Basics to the Hardware step, where it filters the GPU list
• Creating a job became a sequence of sections rather than one form

CLI

Added

• Serve a Managed Inference model across two nodes with a 16-GPU count
• See a reasoning model's thinking in inference chat, and hide it with --disable-think
• Set how much effort a reasoning model spends with --reasoning-effort
• Choose which lifecycle events a job reports with --notify
• Set a lookback window on models healthcheck with --period
• See per-GPU resource specifications and replica node names in jobs detail
• Set a connection timeout on jobs shell with --timeout

Changed

• cosmicac config set replaces admin config set, and needs no administrator credentials

Removed

• API keys as a command-line value, since --api-key takes no value

API

Added

• Send chat_template_kwargs and reasoning_effort on a chat completion, and read the model's thinking in the response
• Read and write observability settings
• Read job-scoped logs
• Read the per-replica GPU count and the replica count on job responses

Deployment

Added

• Run the observability aggregator as part of the deployment
• Pull root disk images to cluster nodes ahead of time with a DaemonSet

Removed

• The built-in log store, replaced by Loki

1.0.1 - 2026-07-17

Stabilisation and completion release following the initial version. Adds Parakeet speech-to-text serving and endpoint health monitoring, and fixes job-list reliability issues found in early operation.

Platform

Web

Added

• Deploy and manage Parakeet speech-to-text endpoints through Managed Inference
• Track Managed Inference endpoint health, latency, and per-replica status from the Model health page
• Replica overview shown separately from metrics statistics on job pages
• Diagnostic bundle collection when submitting feedback

Fixed

• Job list sorting failed with more than 10 jobs and the last job wasn't displayed
• Job statuses now update automatically without refreshing the page
• Failed replicas reported a generic "VMI failed" message instead of the real failure cause

CLI

Added

• Transcribe audio with a Parakeet endpoint

API

Added

• Transcribe audio to text through a Parakeet endpoint

Deployment

Added

• Installation guide for adding model master data to an instance
• Improved seeding of model master data for supported models

1.0.0 - 2026-06-26

Initial release of self-hosted CosmicAC. This entry captures the user-facing capabilities available across the platform and its deployment tooling at 1.0.0.

Platform

Web

Added

• Create and manage GPU Container Jobs from the web interface
• Create and manage Managed Inference Jobs serving vLLM chat models from the web interface
• Create and manage inference API keys from the web interface
• Submit feedback locally with optional webhook delivery

CLI

Added

• Create, list, inspect, restart, and delete GPU Container and Managed Inference Jobs
• Open an interactive shell into a running GPU Container Job
• Chat with a Managed Inference endpoint using streaming responses and multimodal text and video input
• List deployed models and check endpoint health
• Create, manage, and track usage of inference API keys
• Register and manage racks and configure the environment as an administrator
• Connect the CLI to your active deployment

API

Added

• Send chat completions to a Managed Inference endpoint over an OpenAI-compatible API

Deployment

Added

• Deploy the entire self-hosted CosmicAC stack with a single bootstrap command
• Set up model masters to control the models CosmicAC serves and their default parameters
• Upgrade a running deployment to a new image release