Intel / repo

open-webui v0.12.5

Source: github.com
Author: github-actions[bot]
Date: Published Oct 29, 2026

Context

Anyone running Open WebUI, the self-hosted chat interface for local and hosted language models, should treat version 0.12.5 as a security update, even though the release notes do not say what it fixes. The project states that the release contains security fixes that are deliberately left out of the changelog, strongly recommends updating every deployment as soon as possible, and says it will publish the advisories after CVE numbers are assigned. Until those appear on the project's security advisories page, administrators cannot judge which features or configurations were exposed, so upgrading is the practical response.

The rest of the release is mostly administration and reliability work. One fix changes numbers administrators may already rely on: when a model provider repeated its usage report on every streamed chunk, Open WebUI added every report together, so counts for replies and in Admin Analytics were inflated. They now match the provider's final count, so usage figures recorded before the update with such a provider may be higher than the provider's own count. A new Sessions tab in Admin Analytics lists each chat with its tokens, errors and tool calls, but only when the ENABLE_ADMIN_CHAT_ACCESS setting is on, and administrators can now rotate or delete a user's API key from the Edit User dialog. The release also fixes a typing bug for Japanese and other input-method users that version 0.12.4 introduced, where pressing Backspace while composing text could end the composition or duplicate it.

Key takeaways · AI-distilled

Open WebUI 0.12.5 fixes inflated token counts: when a provider re-sent its usage on every streamed chunk, each report was added up. Usage per reply and in Admin Analytics now matches the provider's final count, so earlier analytics may overstate tokens.
Knowledge base API calls that add, update, remove, rename or share now return only the affected files or an empty list instead of the full file list, so they no longer slow down as a knowledge base grows. Clients needing the full list must use the paginated files endpoint.
Open WebUI 0.12.5 makes knowledge base ls, cat and grep commands up to ten times faster on large bases by listing files without loading their full text, and embeds large documents in batches of 1,000 chunks by default to cut peak memory.
In Open WebUI 0.12.5, an unreachable server now fails only that tool call instead of stopping the whole chat, and a foreground that hits the tool-call limit hands back its partial output with the error rather than losing it.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
subagent — A separate agent spawned by another to do one scoped piece of work in its own context, returning only the result.

Read the source github.com