Skip to main content

CubePlex 0.9.4

Features and improvements​

  • Discover and import models from a provider endpoint. Adding a provider lists the models that endpoint exposes, including CLIProxyAPI, vLLM, and Ollama. Review declared, suggested, and conflicting fields, fill in missing limits, and import up to 100 models at a time. Existing model IDs are kept. CLIProxyAPI listings can be filled from the models.dev snapshot shipped with CubePlex. See Discover and import models.
  • Imported models stay disabled until the current configuration passes. Verification checks streaming, a completed tool call, and declared reasoning and image input. Editing the endpoint, key, limits, or calling configuration marks older evidence as verification pending. Models that are already enabled stay available. A disabled imported model needs a passing test of its current configuration before it can be enabled. Model Settings lists only enabled models.
  • Edit calling configuration on the model. The model dialog shows whether a model inherits the provider default or stores its own override. Use provider configuration clears that override. A model with its own override shows a Custom badge.
  • A conversation can use the model default for reasoning. Model default leaves reasoning to the saved endpoint configuration and the model's own default. See Model defaults and available controls.
  • Compaction keeps its summary. CubeLoop 0.15.2 keeps a stored conversation summary when a message's run id, time, or metadata changes after it is stored. When no summary is in effect, the compaction trigger also uses the character estimate, so a long conversation is less likely to resend an over-limit history.

Fixes​

  • Custom providers stay first when adding a provider. Custom OpenAI-compatible and Custom Anthropic-compatible remain ahead of named vendors.
  • A name or slug conflict is reported before the model list. The configuration step saves the provider. Import adds models to that saved provider.
  • A live listing wins when it only disagrees with the catalog. Import uses the listing value, and you can still switch back to the catalog. If required settings are still missing, the message names the model and the fields.
  • The model dialog stays open. Opening a model no longer resets because the omitted provider capability default was a new object on every render.
  • Official DeepSeek Flash discovery uses deepseek-flash. DeepSeek's model list returns deepseek-flash for V4.1 Flash. The official catalog id and display name now match, with display name DeepSeek-V4.1-Flash. The Volcengine catalog entry is unchanged. Stored model rows are not renamed.

Before upgrading​

Upgrade the backend, frontend, and sandbox image together to v0.9.4, or install Helm chart 0.9.4. The normal migration step adds discovery bookkeeping on providers, credentials, and models, and allows a conversation to leave reasoning empty so it uses the model default. Docker Compose backend-migrate and the Kubernetes upgrade job apply these changes. This release does not use the 0.9.0 maintenance flag.

Existing enabled models stay enabled and selectable. A model that already has a test result but no configuration fingerprint shows Verification pending until it is retested. Models that have never been tested stay available. Stored prices, including zeros, are left unchanged. A preset that already points at deepseek/deepseek-v4-flash keeps using that row. New official DeepSeek preset examples use deepseek/deepseek-flash.

If your organization has a saved sandbox policy, update its default image to ghcr.io/cubeplexai/cubeplex-sandbox:v0.9.4. The saved policy takes precedence over the deployment setting. Existing running sandboxes keep their current image until restarted after their tasks finish.

GitHub Release · Full changes since 0.9.3.