Release 0.78.0

March 3, 2026

Caution

v0.78.0 and v0.78.1 are impacted by a critical security issue that allows unauthorized access to an artifact. Do not install or upgrade to this version. Instead, install v0.78.2, which was released March 10, 2026 and includes the fix.

W&B Server 0.78.0 strengthens security and access defaults, adds tools to compare experiments and control costs more efficiently, and more.

For deployments using Weave, Clickhouse v0.25.8+ is required before upgrading to v0.78.

• In Models, create multi-metric plots efficiently with regular expressions, pin important runs, compare experiments against a baseline, and streamline workspaces using enhanced workspace templates.
• Registry includes enhanced search, redesigned cards, clearer and more responsive lineage graphs, and more.
• In Weave, the Trace Analytics dashboard helps you spot bottlenecks and budget drivers from the Project Overview page. Saved views now support customized Traces, Evaluations, and Leaderboards. Weave Playground adds support for LoRA (Low-Rank Adaptation) and new models.
• Inference includes new models.

Support and end of life

• v0.66.0 and below have reached end of life.

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

Security

• A W&B organization or deployment can now have a maximum of 100 Personal API Keys, 100 organization-level API keys per user, and 100 API keys per Service Account.
• On the Enterprise plan, when you invite a member to the organization, they are assigned the Member role by default. On other plans, the default role is Admin.
• When inviting a new user to a team who is not currently a member of the organization, their initial role is now Member.
• The initial role for a user invited to an organization or team now displays in the invitation table.

Features

• Use pinned and baseline runs to keep track of important runs and efficiently evaluate model experiments.

  • Pin up to 6 runs to keep them visible in the workspace and at the top of the runs list. If you have a baseline run, you can pin up to 5 runs because the baseline is implicitly pinned.
  • Specify a baseline run as your reference point for comparisons. The baseline run is always visible in the workspace and at the top of the runs list.

  In the runs table, summary metric deltas show how each run compares to the baseline. This is available by invitation only. Contact [mailto:support@wandb.com] for details.

  In line plots, the baseline appears with visually distinct styling to help with comparison.

• Create panels with multiple metrics in bulk to quickly organize your workspace and spot patterns across large metric sets. For example, you can overlay gradient metrics across model layers or position specific layers side-by-side. Watch a video demonstration.

• Workspace templates have new settings. You can now configure your workspace template to hide empty sections, sort panels in a section alphabetically, or adjust how automatic panel sorting in a section works. See Workspace layout options.

• The Sweeps table now includes a column with the Sweep ID.

• Admins can now customize usage and spend alerts for your organization. View usage and spending and set up email notifications for storage, tracked hours, Weave, Inference, and Training in your organization settings under the new Usage & Alerts tab. See Plan usage.

Registry

• The Registry UI is faster to navigate and easier to scan.

  • Improved type-ahead search, redesigned Registry and Collection cards, and new Registry icons help visually distinguish assets using smart ranking initially, with AI ranking coming soon.

  • Registry icons are initially auto-generated and can be customized by Registry Admins at any time.

    

  • In lineage graphs, when you hover over a node, it is highlighted and other nodes are dimmed.

  • You can now click the name of a collection to quickly view its artifact versions.

  • For convenience, you can now star a registry from its Settings page.

Inference

• The model Moonshot AI Kimi K2.5 is now available. See Model Catalog.

Weave

• The Project Overview page now includes a Trace Analytics dashboard with request counts, latency percentiles (P50/P95/P99), token usage by model, and cost breakdown. Spot which models are burning through your budget, or find the slow requests dragging down your P95. Click a time bucket to jump straight to the relevant traces, to quickly identify the cause of a latency spike.

  

• Weave Playground now supports LoRA, so you can test your fine-tuned LoRA adapters efficiently. Click the Try in Playground button on a LoRA artifact page.

• Claude Opus 4.6 is now available in Weave Playground.

• You can now re-order columns in Traces and Evaluations tables.

  • Drag and drop leaf columns directly in the table or move entire sections around from the column management menu.
  • Persist your changes in a saved view. View a video demonstration.

• You can now save a customized Leaderboard in a saved view.

• You can now evaluate audio and image content in your traces using monitors.

• Aggregated OpenTelemetry token counts now display in the trace detail sidebar for OpenTelemetry traces.

• For models like Gemini that provide thinking tokens, they now display in Token Usage in the Weave UI. Thinking tokens are internal tokens used by LLMs to process complex reasoning steps, calculations, and self-reflection before generating a final answer.

Fixes

• Fixed a bug where a hidden team's reports could be shared.
• Fixed a UI bug where copying text with newlines could truncate the selection.
• When the Sweeps agent claims a preempted run for re-execution, the training script now has up to 60 minutes to start up and initialize the run before the run is sent to a different agent, increased from 60 seconds.
• Fixed a race condition that could prevent a preempted Sweeps run from being enqueued for re-execution.
• Fixed a bug where a Sweep could fail if its name started with a digit.