In this article
LocalAI 5.0.0 moves chats onto the server
LocalAI v5.0.0 reworks the project around chats that run on the server instead of in a browser tab. Around that it ships a desktop app, a file manager, Office document tools, a command sandbox, sharing between users and an add-on library. It also drops the old agent APIs and the liquid-audio backend, so read the upgrade section below before you touch a shared server.
LocalAI runs open models for chat, images, speech and video on your own machines, and its API speaks the OpenAI and Anthropic formats.
Server chats, Files, sandbox and add-ons arrive in 5.0.0
- Server-side chats: sessions keep going after you close the tab, with a message queue, sub-agents, compaction, goals, slash commands and hands-free voice.
- Projects: chats are grouped around one folder, including a memory folder of notes that chats read and you confirm before they are saved.
- Files: folders, upload, preview, search inside documents, archive extraction, a Trash and a History tab that can undo any change made by you or the model.
- Office tools: 11 tools read, edit and create Word, Excel and PowerPoint files, and read and create PDF.
- Sandbox: a bash tool runs Python and shell commands on your files in a KVM or gVisor isolated runtime, with no network by default and each command stored as one undoable change.
- Add-ons: separate programs add tools to a chat, such as Slack and Discord bots, Mail, Google Drive, GitHub and Home Assistant, with install by tag and a single Updates page.
- Sharing: accounts, roles and per-user features, with one Share dialog for folders, saved agents, Knowledge collections and connections.
- Remotes: another LocalAI instance, or an OpenAI or Anthropic compatible endpoint, can be connected and its models used like local ones.
Pages let the assistant build small HTML apps that store their data in your files and open beside the conversation. On the model side, the release adds model search in Explore, an experimental Strata backend, LFM3-Audio speech-to-speech and a forecast API. A first-run Home screen now suggests a model for your hardware.
Only LocalAI's own counts exist so far
The release notes compare against v4.12.1 and count 130 merged PRs, 1,432 non-merge commits, 41 feature PRs and 35 user-facing fix PRs, with more than 80 backends. The model library grew from 2,212 to 2,325 entries, with 134 added and 21 removed.
These are the project's own figures. The notes carry no benchmark scores, and no independent evaluation is cited.
Downloads and breaking changes in LocalAI 5.0.0
The desktop app is an Electron build for macOS on Apple silicon and for Linux on x86-64 and ARM64. It can install, update, start and stop LocalAI, or connect to a server elsewhere, and it reads the same ~/.localai folder. Files are LocalAI.dmg, LocalAI-Launcher-x86_64.AppImage, LocalAI-Launcher-arm64.AppImage, the server binaries local-ai-v5.0.0-linux-amd64 and local-ai-v5.0.0-linux-arm64, and a darwin build for Apple silicon. The older launcher does not update itself to the new app, so download the new file once; models, backends and settings in ~/.localai stay. The sandbox is Linux only in this release.
Check the commands and paths below against the release notes before running anything. Release builds now read the model library (official-models) and the backend library by default, and gallery/index.yaml is a generated copy kept for older releases. Breaking changes for servers:
- Old data: on first start the server converts the old chat page, agents, agent jobs and skills into saved agents, triggers, skills and chats, leaves the old data in place, and keeps converted triggers, MCP connectors and bindings off until you enable them.
- Browser-only chats: upload them with
POST /api/chats/import/browser, and read the conversion report, which names each dropped setting, atGET /api/chats/conversion. - Removed APIs:
/api/agents,/api/agents/skills,/api/agents/git-reposand/api/conversationsnow return 404, so switch to/api/chats/agents,/api/chatsand/api/chats/skills. - Collections and CLI: the collections API moved to
/api/collections, the old path still works with aDeprecation: trueheader, and thelocal-ai agentcommand is gone. - Settings:
LOCALAI_AGENT_POOL_*becameLOCALAI_COLLECTIONS_*, andLOCALAI_DISABLE_AGENTSsplit intoLOCALAI_DISABLE_CHATSandLOCALAI_DISABLE_COLLECTIONS; old names still work with a warning, and some agent pool options are ignored. - Studio and Talk: both are now add-ons with nothing preinstalled, an administrator installs them, and users need the addons feature, which the upgrade does not grant.
- Permissions: on servers with accounts, saved agents, triggers, connectors, pages and add-ons are off for new users until granted, and the skills and chat_history features no longer exist as permissions.
- Audio: configs using
backend: liquid-audiostop working; switch to anlfm3-audio-2bentry on llama-cpp withpipeline.any_to_any: true, while LFM2-Audio and LFM2.5-Audio have no replacement. - Office tools: the creation tools now take Markdown or CSV, and the old
blocksandsheetsarguments are rejected with migration guidance.
llama-cpp models that use realtime_audio must now also set pipeline.any_to_any: true. /v1/chat/completions is unchanged, so clients on the OpenAI-style chat endpoint should not need edits.
Who should plan a LocalAI 5.0.0 upgrade
If you run LocalAI for other people, treat this as a migration, not a routine update. Try it on a copy first, check which scripts and clients call the removed agent paths, and decide who gets the new features. If you run it alone on one machine, the main work is reading the conversion report and re-enabling triggers and connectors you still use.
What the LocalAI notes leave out
The release text we reviewed gives no release date, licence or price, and no hardware needs beyond KVM or gVisor for the sandbox. It was also cut off, so later sections and any further downloads are not covered here; the full release notes may list more.
Sources
Image: GDJ, CC0 1.0, via Wikimedia Commons
This story was produced with AI assistance from the sources listed above and fact-checked against them before publication. Read our editorial methodology and the corrections log. Spotted an error? Email [email protected].





Discussion