Skip to content

v5.0.0

Latest

Choose a tag to compare

@mudler mudler released this 18 Nov 17:00
· 3 commits to master since this release

site-release-1920x1080-v3

Run open models for chat, images, speech and video on your own machines. The API speaks the OpenAI and Anthropic formats. Your team gets chats, files and documents on top, and the assistant can write and run code. 5.0 adds a desktop app, Projects, a Files page, Office documents, a Python sandbox, team sharing and an add-on library, all built on chats that now run on the server.


Highlights:

  • 🖥️ Desktop app: a new app for macOS (Apple silicon) and Linux (x86-64 and ARM64) installs, updates, starts and stops LocalAI, or connects to a LocalAI server somewhere else.
  • 🗂️ Projects: group chats around one folder, with notes the chats read and that you confirm before they are saved.
  • 💬 Chats on the server: a chat keeps working when you close the tab. Steer it while it works, hand work to sub-agents, set a /goal, or talk to it hands-free with Live.
  • 📁 Files: upload, preview, search, extract archives, Trash, and a History tab that undoes any change, yours or the model's.
  • 📄 Office documents from a chat: the Simple Office add-on reads, edits and creates Word, Excel and PowerPoint files, and reads and creates PDF.
  • 🐍 Sandbox: run Python and shell commands on your files in an isolated runtime with no network by default. Linux only in this release.
  • 👥 Team and sharing: accounts and roles, plus sharing of folders, saved agents, Knowledge collections and connections with one person or with everyone.
  • 🧩 Add-ons: a library of separate programs that add tools to a chat, from Slack and Discord bots to Mail, Google Drive, GitHub and Home Assistant.
  • 🔌 Remotes: connect another LocalAI instance, or an OpenAI or Anthropic compatible endpoint, and use its models like your own.
  • 🎨 New look and a better first run: a new theme, logo and mascot, and Home recommends a model for your hardware.

📊 This release in numbers

Pull requests merged 130
Commits 1,432 (non-merge), most of them direct commits
Files changed 7,310
Feature PRs 41
User-facing fix PRs 35
Model library 2,212 to 2,325 entries (+134 added, 21 removed)
Backends more than 80

Counts compare v4.12.1 with the release-preparation snapshot. Most of 5.0 landed as direct commits, so the sections below cite commits and docs pages as well as PRs.


📌 TL;DR

Area Summary
🖥️ Desktop app Electron app for macOS (Apple silicon) and Linux AppImages. It replaces the older launcher and reads the same ~/.localai folder.
🗂️ Projects Chats grouped around one folder, with a memory/ folder of notes, defaults for new chats, and agents made from a project.
💬 Chats Chats run on the server, with steering, a message queue, sub-agents, compaction, goals, slash commands and Live voice. /v1/chat/completions is unchanged.
📁 Files Files page with folders, upload, preview, search inside documents, archive extraction, Trash and undoable history.
📄 Office Simple Office add-on: 11 tools for Word, Excel, PowerPoint and PDF, and PowerPoint decks built on your own template.
🐍 Sandbox bash tool in a KVM or gVisor sandbox with your files mounted. Each command is one undoable change. Linux only.
👥 Team Roles, per-user features, and one Share dialog for folders, agents, collections and connections.
📑 Pages The assistant writes small HTML pages that keep their data in your files and open beside the chat.
🧩 Add-ons Add-on library with install by tag, upgrade review, one Updates page, OAuth login for add-ons, and Slack and Discord chat apps.
🔎 Models Remotes, model search in Explore, the model library, an experimental Strata backend, LFM3-Audio speech-to-speech and a forecast API.
🎨 UI New theme, logo and mascot, a first-run model recommendation, a redesigned Nodes page.

⚠️ Upgrade notes

5.0 replaces the chat page and the agent runtime, and moves Studio and Talk to add-ons. Read this before you upgrade a server that other people use.

  • Chats replace the old chat page, agents, agent jobs and skills. On the first start the server converts what the old features left on disk and in the database into saved agents, triggers, skills and chats. It does not change or delete the old data. Converted triggers, MCP connectors and connector bindings are switched off until you turn them on. Conversations kept only in a browser were never on the server: upload them with POST /api/chats/import/browser. GET /api/chats/conversion returns the report, and each dropped setting is named in it. See Chats: moving from the old chat, agents and skills. #12918
  • Removed APIs. /api/agents (agents, actions, SSE), /api/agents/skills, /api/agents/git-repos and /api/conversations answer 404. Use /api/chats/agents, /api/chats and /api/chats/skills. The collections API moved to /api/collections; the old path still works with a Deprecation: true header. The command local-ai agent is gone.
  • Renamed and ignored options. The LOCALAI_AGENT_POOL_* collection settings are now LOCALAI_COLLECTIONS_*, and LOCALAI_DISABLE_AGENTS is split into LOCALAI_DISABLE_CHATS and LOCALAI_DISABLE_COLLECTIONS. The old names still work and log a deprecation warning. Some agent pool options are now ignored; the migration page lists them.
  • Studio and Talk are add-ons. Nothing is preinstalled. An administrator installs them from Add-ons, and users need the addons feature, which the upgrade does not grant. See Upgrading: Studio and Talk are add-ons now.
  • Permissions. On a server with accounts, saved agents, triggers, connectors, pages and add-ons are off by default for new users until an administrator grants them. The skills and chat_history features are gone from the user permissions.
  • liquid-audio is removed. Configs with backend: liquid-audio stop working. Use an lfm3-audio-2b entry on llama-cpp with pipeline.any_to_any: true. LFM2-Audio and LFM2.5-Audio have no replacement. llama-cpp models with realtime_audio must now set pipeline.any_to_any: true. #12958
  • Simple Office creation tools take Markdown or CSV. The old blocks (DOCX, PDF) and sheets (XLSX) arguments are rejected with migration guidance. #13029
  • The desktop app replaces the older launcher. The older launcher does not update itself to the new app: download the new file once. Your models, backends and settings in ~/.localai stay. See Move from the older launcher. Commits 0a72543, 822a7bc.
  • Models and backends come from the libraries. Release builds read the model library (official-models) and the backend library by default. gallery/index.yaml is now a generated copy, kept for older releases. See Model library.
  • Add-ons follow the release. A release build reads the add-on index attached to its own release. A source build reads the development channel.

📦 Downloads

Platform File
macOS, Apple silicon (desktop app) LocalAI.dmg
Linux x86-64 (desktop app) LocalAI-Launcher-x86_64.AppImage
Linux ARM64 (desktop app) LocalAI-Launcher-arm64.AppImage
Linux x86-64 (server binary) local-ai-v5.0.0-linux-amd64
Linux ARM64 (server binary) local-ai-v5.0.0-linux-arm64
macOS, Apple silicon (server binary) local-ai-v5.0.0-darwin-arm64

Each file has a .sha256 file next to it. There is no macOS build for Intel: use a container there. The AppImage needs FUSE 2; on Ubuntu 24.04 and later install libfuse2t64. See the desktop app guide.

To install a server instead:

curl -sSL https://localai.io/install.sh | sh

or with a container:

docker run -p 8080:8080 --name local-ai -ti localai/localai:latest

🚀 New Features & Major Enhancements

🖥️ The desktop app

LocalAI has a desktop app for macOS and Linux, built on Electron. It installs, updates, starts and stops the LocalAI server on your computer, and shows the web interface in its own window with a status bar for the server. On the first start you choose to run LocalAI on this computer or to connect to a server that runs somewhere else, and you can keep several servers. It has pages for logs, settings, and a Connect page that shows how other devices and apps reach the server.

The DMG and the app are signed with an Apple Developer ID and notarized. On Ubuntu 23.10 and later, the AppImage starts with the Chromium sandbox off unless user namespaces are allowed; the status bar says so, and the app can open the web interface in your browser instead. The app replaces the older Go launcher.

02-desktop-app-first-run
On first start, the desktop app asks where LocalAI should run: on this computer or on a LocalAI server.

03-desktop-app-models
The Models page lists the models that fit this computer, with the best fit first.

🔗 Commits: Electron launcher, remove the Go launcher, DMG build. Docs: Desktop app.

🗂️ Projects

A project groups chats around one folder in your own files. The folder holds a files/ folder for the material and a memory/ folder of notes that the project's chats read at each turn. The model can propose a note, and you confirm it before it is saved. The Tools tab sets defaults for new chats: skills, add-ons, MCP servers, connectors and collections. A chat has one home project and can read up to eight others. You can make a saved agent from a project and share the project folder.

A project organises chats. It is not an access boundary: a chat outside the project can still read the project's files from your home folder.

projects
A project page: recent chats, the project memory and the tabs for files, agents and tool defaults.

🔗 Commit: add Projects. Docs: Projects.

💬 Chats that run on the server

Chats are now kept and run by the server, so a chat keeps working when you close the tab. A message you send while the model works waits in a queue, can be edited or discarded, and reaches the model at its next step. The model can start sub-agents in the foreground or the background (up to four in the background), and long chats are compacted. The OpenAI-compatible /v1/chat/completions endpoint is unchanged.

  • Goals. /goal <text> keeps a chat working toward an objective over many turns. It pauses on Stop and at its limits: 20 turns and 400 steps by default (chats_goal_max_turns, chats_goal_max_steps).
  • Slash commands. /compact, /clear, /loop, /skill, and your own commands.
  • Live. Talk to a chat hands-free from the chat or from Home. Live needs a microphone, a speech-to-text model and a voice model; an administrator installs small ones with one click from the setup sheet. Your audio stays on your LocalAI server.
  • Memory and change proposals. Every chat can read the index of your notes with memory_read. A model asks to change a note or an agent profile with propose_change, and nothing is written until you confirm. A confirmed change can be undone.
  • MCP connector tools behind the tool search. With chats_tool_search on, MCP connector tools are found on demand instead of being sent on every turn.
  • Send a message from a chat. With send_message, a chat and its loops can post to an approved channel of one of your messaging connections, for example to report a result. One chat sends at most 5 messages in 10 minutes.
  • Context size for the chat agent. For a chat model with no context_size, the loader aims at 32,768 tokens and steps down to what fits, never below 16,384 while that fits. A banner says when a model's window is too small.

07-agents-create
Create a saved agent in four steps: purpose, abilities, where it works, and a review.

🔗 PRs: #12918, #13064, #13066, #13069, #13163, #13153, #13199. Commits: Live, Live from Home, user slash commands, send_message. Docs: Goals, Live.

📁 Files

The new Files page holds your files. Upload files and folders, drag rows between folders, preview a document and download it. Search covers names and the text inside PDF, Word, Excel and PowerPoint files. Extract here unpacks ZIP and TAR archives; 7z and RAR stay off until an administrator turns them on. Deleting moves a file to the Trash, and the History tab undoes a change.

A chat uses the same files. Attach a file to a message, or let the model list, read, search, write and edit your files. What the model changes goes into the same history, so you can undo it. A saved agent can be limited to chosen folders, each as read or write. Distributed deployments without S3 can keep the files on a shared filesystem.

06-files-history-undo
Every change is listed in History with an Undo, including the commands a chat ran in the Sandbox.

🔗 PRs: #13009, #13012. Commits: extract an archive, managed folders. Docs: Files and shared folders.

📄 Office documents from a chat

The Simple Office add-on has 11 tools: read, edit and create for Word, Excel and PowerPoint, and read and create for PDF. Creation takes Markdown (Word, PDF) or CSV (Excel). A PowerPoint deck can be built on your own .pptx or .potx template and keeps its masters, layouts, theme colours and fonts. When a file name is taken, the new file is saved next to the old one. Limits: no PDF editing, the PDF writer takes Latin-1 text only, and layouts are basic. For charts or exact layout, use the Sandbox.

04-office-preview
Preview a spreadsheet in Files, next to the Word, PDF and PowerPoint files of the same plan.

🔗 PRs: #13029, #13227. Docs: Office.

🐍 Sandbox: run Python on your files

The Sandbox add-on gives a chat a bash tool that runs in an isolated runtime with your files mounted. Ask for a chart from a CSV file, and the model runs a Python script and saves the image in your files. What a command changed is saved as one change that you can undo. The add-on tries KVM (Cloud Hypervisor) first, then gVisor. The base profile has Python with pandas, matplotlib, openpyxl, python-docx, python-pptx, pypdf and Pillow; the media profile adds ffmpeg, sox and ImageMagick.

A sandbox has no network device. HTTP and HTTPS go through a proxy, and the default is Off, so pip install fails until an administrator allows package registries or public sites. A chat can ask for a wider setting, but never for more than the administrator allows. The Sandbox add-on runs on Linux (amd64 and arm64) in this release.

05-sandbox-chart
The chart comes back as a file in the chat, created by the assistant.

🔗 PRs: #13038, #13073. Docs: Sandbox.

👥 Team and sharing

With accounts on, there are administrator and user roles, and each person has their own files, chats and saved agents. One Share dialog shares a folder of your home, a saved agent, a Knowledge collection or a connection with one person or with everyone. Revoking access stops the work that the new role no longer allows. A shared saved agent runs as the person who uses it, with that person's folders, memory and connections. An administrator can also make shared folders and choose their members.

Only a person can share, from the browser. No model, agent or add-on can share anything or take access away.

🔗 PRs: #13130, #13009. Docs: Sharing.

📑 Pages

Ask the assistant for a small HTML page, such as a task board, and it writes one that keeps its data in your files. Pages open in a dock beside the chat, and a trigger can keep a page's data current. For users on a server with accounts, pages are off until an administrator switches them on.

10-pages
A task board page made by the assistant, open beside the chat.

🔗 PR: #13127. Commits: pages beside the chat. Docs: Pages and Home.

🧩 Add-ons and the add-on library

An add-on is a separate program that gives LocalAI new tools. An administrator installs it from the library, and its tools appear in the chat's Tools menu. The library includes Studio and Talk, Voice, Face, Fine-tuning, Quantization, Tables, Simple Office, Sandbox, Web, Calendar, Contacts, Mail (IMAP and SMTP), Gmail, Google Drive, GitHub, SSH, Home Assistant, Jellyfin and PiKVM.

  • Chat apps. Slack and Discord add-ons let people talk to a saved agent from those apps. A binding answers nobody until you add allowed senders, and it can limit the kinds of file it reads.
  • OAuth for add-ons. An add-on can ask a person to log in to an external service with OAuth 2.0; the host runs the login after an administrator approves it.
  • Write modes. Mail, Gmail, Google Drive and PiKVM offer Read only, Asks you first and Allow all.
  • Updates in one place. An Updates page covers backends and add-ons, with a review step before anything runs and inbox notifications.
  • Install and run. Development add-ons install by tag and are verified against the resolved digest. The add-on owner can run as its own workload in a cluster (--addons-role).

An add-on runs as the operating system user that runs LocalAI, so install only add-ons you trust. Only the Sandbox isolates what it runs.

08-addons-tables
The Tables add-on: a Library database with a Books table that you can filter, export and share.

09-addons-library
Find add-ons in the library and install them from the card.

🔗 PRs: #12858, #12867, #12875, #12883, #12898, #12948, #13034, #13077, #13097, #13104, #13128, #13154, #13167, #13173. Docs: Add-ons.

🔌 Remotes

Connect another LocalAI instance, or an OpenAI or Anthropic compatible endpoint, from the first-run guide or from the Models page. Each model the remote offers becomes an ordinary model config, so it works in chats, the API and failover. The credential is sealed at rest and write-only in the API. The admin API is under /api/remotes.

🔗 PR: #13235. Docs: Remotes.

🔎 Models, backends and engines

  • Model search. The Explore tab searches Hugging Face, with the size of each file and whether it fits your machine. Gated Hugging Face repositories are not offered.
  • Strata (experimental). A backend for the Strata engine runs Qwen3.8-Flash-Next and DeepSeek V4.1-Flash on supported GPUs, with gallery entries that build the engine's pack on first load. #13125, #13131, #13132
  • LFM3-Audio speech-to-speech. LFM3-Audio-2B runs on llama-cpp in the Realtime API (WebSocket and WebRTC). #12958
  • Forecasting. A timestral-cpp backend and a POST /v1/forecast endpoint return quantile forecasts for time series, on CPU. #12885
  • Vision alias. An alias can name a vision_alias model that serves its requests with images. #13171
  • llama.cpp. New moe_cache_cpu_share option, and fixes for reasoning_budget and mlock. #13155, #13157
  • Backend operations. --max-concurrent-backend-operations lets installs, upgrades and deletes run in parallel (default 1). #12888
  • Realtime. Transcription-only sessions announce the speaker of each turn. #12871

🔗 Commits: model search API, search from Explore.

🎨 New look and first run

The web UI has a new theme, logo and mascot, in light and dark. On a fresh install, Home recommends a model for the hardware it detects, and the first-run guide asks whether your models run on this machine or on another server. The Nodes page is redesigned with a fleet card, grouped rows and an inspector. iOS shows the logo when you add LocalAI to the home screen.

x09-look-before-after-1600x900
Home in 4.12 on the left, and in 5.0 on the right.

🔗 PRs: #13187, #13199, #13231, #13121, #12917.


🐛 Bug Fixes (recap)

  • fix(distributed): keep one idle replica when scaling down, and keep healthy replicas when one fails. #13201, #13233
  • fix(chats): keep the system prompt stable across chats so the prompt cache hits, and start every shell in the chat's work folder. #13153, #13212
  • fix(llama-cpp): fix reasoning_budget and mlock, add missing options, harden raw flags; keep an existing image data URL. #13157, #12921
  • fix(downloader): parse the Hugging Face @revision suffix without panicking. #13179
  • fix(oci): fall back to the OS temp folder when the staging folder is not writable. #13186
  • fix(vram): do not cache failed probes. #13203
  • fix(realtime): never offer deprecated IPv6 addresses as WebRTC ICE candidates. #13172
  • fix(cloud-proxy): collapse nullable type unions in OpenAI tool schemas. #13165
  • fix(proxy): propagate model env and accept direct keys. #12940
  • fix(trace): stop trace writes into the data path on shutdown. #13161
  • fix(core): correct AMD discovery and MCP and P2P handling. #13044
  • fix(strata): preserve image markers in tokenizer messages. #13254
  • fix(mlx-vlm): bump the mlx-vlm pin to v0.6.12. #12818
  • fix(ui): show GPU memory use on the Settings page, no "No models loaded" for non-admins, keep the Models page tabs usable. #13162, #13166, #12899
  • fix(addons), fix(connectors), fix(studio): owner settings, upgrades, discovery and connector delivery fixes. #12850, #12878, #12943, #13026, #13030, #13033, #13035, #13062, #13078, #13127, #13129
  • fix(kokoros): implement Forecast as unsupported. #13134
  • fix(gallery): installing a recommended model with no variants from Home no longer fails with "pinned model variant not found". #13277
  • fix(timestral-cpp): build on arm64 (Linux and macOS). #13276

🧠 Models

The model library grows from 2,212 to 2,325 entries: 134 added, 21 removed (the liquid-audio entries among them). Additions include the Strata GGUF sets (Qwen3.8-Flash-Next and DeepSeek V4.1-Flash), Phonon-2.5 and Fluid Speech Pro on parakeet-cpp, Flint, MiniCPM-OCR 1.0, Nanbeige4.3, Gemma 4 and Qwen3.5 Artemis variants, Ternary Bonsai 2, and BGE English embedding variants. Check each model card for its licence and hardware needs.

🔗 PRs: #12854, #12896, #12994, #13131, #13132, #13253, #13267.


👒 Dependencies

The range has 104 engine update commits, most made directly by the update bot, and 19 dependency PRs.

Source Updates
ikawrakow/ik_llama.cpp 15
0xShug0/audio.cpp 15
ggml-org/llama.cpp 14
leejet/stable-diffusion.cpp 12
mudler/vllm.cpp 11
CrispStrobe/CrispASR 9
NVIDIA/NeMo-Speech.cpp 5
ggml-org/whisper.cpp 5
ServeurpersoCom/omnivoice.cpp 3
PrismML-Eng/llama.cpp 3
Niko1221/Strata 3
antirez/ds4 3
Four other sources 6

📖 Documentation

The docs are regrouped into nine sections with a new brand, website and README (#13178). New pages cover the desktop app, Projects, Files, chats and goals, Live, the migration from the old chat and agents, Remotes, the add-on catalogue, and connecting coding agents to the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages APIs. The product lede names coding (#13275). The 5.0 blog post and the home page download buttons are in #13229 and #13271.


🐞 Reporting problems

5.0 carries a lot of new work. If something breaks after the upgrade, open an issue or ask in the Discord.


🙌 New Contributors

Thanks to the contributors making their first merged PRs:

Thanks also to @walcz-de for the chat, connector and Simple Office work in this release.

Full Changelog: v4.12.1...v5.0.0

What's Changed

Breaking Changes 🛠

  • feat(realtime): LFM3-Audio speech-to-speech on llama.cpp, remove liquid-audio by @mudler-agent in #12958
  • feat(chats): server-side chats with steering, sub-agents and compaction by @mudler-agent in #12918

Bug fixes 🐛

Exciting New Features 🎉

🧠 Models

  • feat(gallery): add flint and bump llama.cpp to a37cc461 by @mudler-agent in #12854
  • feat(gallery): add Phonon-2.5 parakeet-cpp models by @mudler-agent in #12896
  • fix(model-gallery): point pinned LFM3-Audio entries at main by @mudler-agent in #12994
  • chore(model-gallery): ⬆️ update checksum by @localai-update-bot[bot] in #13253
  • chore(model-gallery): ⬆️ update checksum by @localai-update-bot[bot] in #13267

📖 Documentation and examples

👒 Dependencies

  • chore(deps): bump actions/setup-node from 6 to 7 by @dependabot[bot] in #12837
  • chore(deps): bump docker/login-action from 3 to 4 by @dependabot[bot] in #12836
  • chore(deps): bump grpcio from 1.84.0 to 1.85.0 in /backend/python/rerankers by @dependabot[bot] in #12963
  • chore(deps): bump docs/themes/hugo-theme-relearn from 657eb0c to 53bcfee by @dependabot[bot] in #12962
  • chore(deps): bump grpcio from 1.84.0 to 1.85.0 in /backend/python/common/template by @dependabot[bot] in #12961
  • chore(deps): bump torch from 2.14.0+xpu to 2.15.0+xpu in /backend/python/common/template by @dependabot[bot] in #12960
  • chore: ⬆️ Update ggml-org/whisper.cpp to 5922450fbce1ed9d1648babd0d5ac4520055b436 by @localai-update-bot[bot] in #13250
  • chore: ⬆️ Update ikawrakow/ik_llama.cpp to fcf7b364786ec621d0c53eae39223efc5492b8c5 by @localai-update-bot[bot] in #13249
  • chore: ⬆️ Update CrispStrobe/CrispASR to 9ddae4590d427f27e0ca04694864eed66bcb530e by @localai-update-bot[bot] in #13247
  • chore: ⬆️ Update 0xShug0/audio.cpp to 89bccbea607d3bb35cc6aff17476ad97e472aa15 by @localai-update-bot[bot] in #13246
  • chore: ⬆️ Update ServeurpersoCom/omnivoice.cpp to a5bd24f6b675335d1c9c0e945793d0de9fabe9a4 by @localai-update-bot[bot] in #13251
  • chore: ⬆️ Update ggml-org/llama.cpp to fc504f91d8cfdfcd9b0fedcf14600fd2b0e63420 by @localai-update-bot[bot] in #13245
  • chore: ⬆️ Update ggml-org/whisper.cpp to aba02defd6bb6ff5de63721c21e3456645c153e2 by @localai-update-bot[bot] in #13266
  • chore: ⬆️ Update ikawrakow/ik_llama.cpp to a6383da414f74c944999e7b9312a5a5d76191f6b by @localai-update-bot[bot] in #13262
  • chore: ⬆️ Update 0xShug0/audio.cpp to 2872e36e992c8dd98fcc47de094bd703e1ff6dab by @localai-update-bot[bot] in #13264
  • chore: ⬆️ Update ggml-org/llama.cpp to d998c676ff463b1b8ad94ad7c0b905e281422fad by @localai-update-bot[bot] in #13261
  • chore: ⬆️ Update Niko1221/Strata to v0.2.15 (c77572806996349b49995e4f98d41b5294d36a21) by @localai-update-bot[bot] in #13265
  • chore: ⬆️ Update ServeurpersoCom/omnivoice.cpp to a5bd24f6b675335d1c9c0e945793d0de9fabe9a4 by @localai-update-bot[bot] in #13263
  • chore: ⬆️ Update leejet/stable-diffusion.cpp to 532d2c15c430a785a712b2d94281f04e733008a4 by @localai-update-bot[bot] in #13248

Other Changes

New Contributors

Full Changelog: v4.12.0...v5.0.0