0.74.x

  4 minute read  

W&B v0.74 delivers SCIM API improvements for enterprise identity management, including automated group sync with Okta, streamlined user role configuration, and more. In Dedicated Cloud, a powerful new full-screen image browser helps you inspect and compare images within and between runs. Weave enhancements include in-Playground LLM judge evaluations, better visibility into multi-turn Thread tool calls, DSPy improvements, and a new integration with the Verifiers library for Reinforcement Learning workflows. W&B Inference adds support for the Z.AI GLM v4.5 hybrid reasoning model.

Support and end of life

• W&B Server v0.60 and below reached end of life on October 2, 2025.
• W&B Server v0.61 is scheduled to reach end of life on October 28, 2025.

A W&B Server release is supported for 12 months from its initial release date. As a reminder, customers using Self-managed are responsible to upgrade to a supported release in time to maintain support.

Refer to Release policies and processes. For assistance or questions, contact support.

Features

• The SCIM API has been enhanced:
  • Provisioning and deprovisioning operations are now more compatible with Okta.
  • In Multi-tenant Cloud deployments, you can now configure the organization role and teams when you create a user. This aligns with the endpoint’s behavior in Dedicated Cloud and Self-Managed.
  • In Dedicated Cloud and Self-Managed deployments, you can now update a user’s username or email address. See Update user email and Update user display name. Not applicable to Multi-tenant Cloud deployments.
  • You can now search and retrieve a user by email and username.
  • Okta groups are now synced automatically using ETags. For details about SCIM groups, see Group resource.
• When exporting the list of users in your organization, the CSV file now includes columns for Models seats and Weave access.
• In W&B Dedicated Cloud, the new full-screen image browser simplifies inspection and comparison of logged images.
  • True full-screen view. Inspect image details more easily with edge-to-edge full-screen viewing. No more cramped panels.
  • Step slider in full-screen view. Move through steps efficiently with an embedded slider to see how outputs evolve.
  • Quickly toggle between runs using the keyboard. In full-screen view, use the keyboard left / right arrow keys to move between runs, and use Cmd + left / right arrow (macOS) or Ctrl + left / right arrow (Windows / Linux) to move between steps of the same image.

Weave

• From the Evaluations sidebar tab, you can now update prompt values and run LLM judge evaluations to test the changes. Try it out in the Playground.
• You can now integrate Weave into your Reinforcement Learning (RL) training workflows and view the traces and evaluations captured during training in the W&B Workspace, together with run metrics panels.
• Weave now includes an integration with Verifiers, a library of modular components for creating Reinforcement Learning (RL) environments and training LLM agents. Environments built with Verifiers can serve as LLM evaluations, synthetic data pipelines, agent harnesses for OpenAI-compatible endpoints, and RL training. Learn more.
• From the Threads sidebar tab, you can now view tool calls for a multi-turn Thread together with its message history.
• A Team’s Project page now shows the number of Weave traces for each project together with data about the project’s runs.
• The DSPy integration has been improved. For details, see wandb/weave #5184.
• If your Weave code calls wandb.init(), it no longer needs to explicitly call weave.init().
• New convenience method delete_all_object_versions.
• The output from an OpenAI streaming endpoint call now includes a new field, time_to_first_token, expressed in milliseconds.

Inference

• W&B Inference adds support for the Z.AI GLM hybrid reasoning model. v4.5 is supported.

Fixes

• Fixed display problems in bounding boxes with captions.
• When grouping runs in the Runs Table, you can now expand the null group to see the runs with null values for the grouping key.
• Increased running average smoothing parameter maximum to 1000.
• Fixed a bug where clicking a run name could open an invalid URL if a runset included runs from multiple projects.
• Fixed text display issues in dark mode.
• Fixed a permission bug when viewing a public report that included Query Panels.
• Fixed a UI bug that could prevent code from being copied from a report.
• Fixed a bug in the order of min and max values in a line plot tooltip.
• Fixed a bug that could prevent a grouped plot from rendering if the X-axis key was a substring of the Y-axis key.