🚀 New Features
- MLX installs on demand — the app now installs and activates the MLX engine through the core when you first use an MLX model instead of shipping it pre-bundled.
- Audios studio (dev builds) — a new Audios prototype (text-to-speech, text-to-music, sound effects) is available in development builds (no engine yet).
- Attachment previews and video attachments — composer and sent messages open previews for images, PDFs, audio, video and common document types, and you can attach video files (models cannot read video contents).
- Help → Chat with us — a built‑in support chat sheet lets you message the Atomic Chat team; the relay forwards replies and can include optional diagnostics.
- Native Linux packages and repo — deb, rpm and AUR packages are produced and a signed apt/dnf repository is available for Linux installs.
- Configure server engine from the picker — a gear beside the chosen server engine opens that engine’s settings in a dialog so you can configure engine/model directly.
- Onboarding model-ready email reminders — the onboarding flow can deliver email reminders when a model is ready.
🔧 Improvements & Fixes
- Downloads resume correctly — resumed downloads keep files already finished (sha256-checked) and reopen low-throughput connections instead of re-downloading finished bytes.
- Engine checks retry and show reasons — manifest and engine build checks retry automatically and report plain error reasons (timeout, network, HTTP status) with a Try again button.
- Corner notifications stacked and auto-fold — corner cards are now a single stacked column and fold into a pill after 3s to avoid overlapping the download panel.
- Tray shows runtime and opens picker — the macOS tray now lists the loaded model and its runtime; clicking opens the app and opens the Model/Engine picker.
- Server cards run agents again — Server cards now walk Install → Configure → Connect and properly open and connect agents from the Server UI.
- Codex runs in its own profile — Codex is launched with a dedicated atomic-chat profile and shared config.toml edits are repaired to avoid breaking other Codex users.
- First-run flow polish — purpose cards simplified, the wait screen shows Atomic Mines while the first model downloads, and the onboarding run persists and prioritizes chat model downloads.
- Voice & media fixes — Stop works while voice models are starting, cancelled voice downloads are not reported as failures, and attached media survive slow core conditions with clearer ENGINE_STOPPED messaging.
- Activity, drafts and projects — Activity logs requests immediately (including timings), drafts persist across relaunch, and project file removal asks for confirmation.
- Integration, support and analytics regressions addressed, plus chat and download-panel fixes.
- 10+ additional stability and UI fixes across engines, installs and the Hub.
⚠️ Breaking Changes
- Installer payload changed — installers now ship only the upstream llama.cpp build; MLX, TurboQuant (llamacpp fork) and jan-cli are no longer bundled and are installed on demand, which reduces download size but requires on‑demand engine downloads for those backends.
- Linux compatibility note — the bundled upstream llama.cpp requires glibc 2.38, so the new Linux packages target Debian 13 / Ubuntu 24.04 or newer.
🙏 Contributors
Thanks to @Ooooze, @Vect0rM for their contributions to this release!