Highlights

AMD ROCm/HIP Backend (Experimental, Linux)

Jan now detects AMD GPUs and offers a HIP (ROCm) llama.cpp backend on Linux. The new hip backend is wired through detection, naming, prioritization (ranked above Vulkan when VRAM is sufficient), download/extract, and dependency verification. Both Jan's own asset format and the official llama.cpp ROCm asset names are folded onto a single canonical backend id, so listing, update checks, and migration treat every HIP variant as one.

Experimental: The entire ROCm/HIP backend is experimental in this release. Only Linux is shipped and supported — Windows ROCm/HIP is not shipped. If you run into problems, please report them on GitHub (opens in a new tab) so we can stabilize AMD GPU support faster.

Branded Startup Loader & Faster First Paint

Startup was reworked so the window paints immediately instead of waiting on extension initialization:

• A branded loader with phased status messages replaces the blank boot screen, with a watchdog against a hung blank window.
• The initial app update check, the llama.cpp backend dependency verification, and the MCP setup are all deferred until after first paint / browser idle, so a cold start is no longer blocked on network probes and version checks.
• The dark theme is applied synchronously on launch (carried over and hardened), and the loader logo/caption spacing was tightened and centered.

Pause & Resume Model Downloads

Model downloads can now be paused and resumed directly from the Hub panel, so a large GGUF pull no longer has to restart from zero after an interruption.

Safer Default Context Size

When no context size is set and auto-fit is off, the llama.cpp router now defaults ctx-size to 8192 instead of letting llama.cpp load the model's full trained context (often 128k+), which could OOM on large-context models. User and per-model values still override this default.

Automatic Update Check Toggle

Settings gained a toggle for the automatic app update check. When disabled, Jan no longer hits the network for updates behind the scenes; the manual check still works on demand.

Provider & Key Management

• API key inputs now have copy and reveal buttons (a new secret-input component).
• Transport-level request failures (connection refused, DNS, TLS) now surface a friendly, actionable message instead of a raw error.
• MLX models are listed in the Local API Server's default model picker.

Bug Fixes

• Sampler settings now persist across restart. Thread updates previously rebuilt the assistant from a field whitelist that dropped parameters, so sampler overrides (temperature, top_p, top_k, etc.) set in the composer were lost on relaunch. The assistant is now persisted intact.
• MLX commands fixed on macOS. The MLX capability was moved into its own macOS-scoped file during the macOS-only build split but never added to the security allowlist, causing a "not allowed by ACL" error when loading MLX models. The capability is now registered.
• Attachment size limit is honored. The document parser hardcoded a 20MB cap regardless of the configured "max file size" setting, so a file under the user's limit but over 20MB (e.g. 21.5MB with a 50MB setting) was rejected as too large. The parser backstop now follows the setting's documented maximum, and the frontend remains the source of truth for the configured value.
• Notification toasts are visible and theme-aware again: the previously page-colored (invisible) card is themed via popover tokens and the active theme with an elevation shadow, and neutral toasts pick up the Appearance accent color.
• Toast styling is restored on Windows/WebView2. Sonner's CSS is shipped via the linked app bundle, and the accent tint moved from an inline style prop to a CSS class so WebView2's hashed style-src CSP no longer drops it.
• CSP no longer blocks runtime inline styles: Tauri's automatic nonce injection on style-src is disabled so 'unsafe-inline' stays effective, fixing the wave of "inline style violates CSP" errors (Google Fonts, dropdowns, etc.).
• The thread's selected agent is now authoritative, so switching the assistant mid-chat works as expected.
• Prompt-too-big errors are routed to the "Increase Context Size" banner instead of a generic failure.
• Crashed models are recovered and mid-request crashes are surfaced rather than hanging silently.
• The router is started before the backend update check is deferred, avoiding a race where the engine wasn't ready.
• requestIdleCallback is bound to window to avoid an "Illegal invocation" error in the updater.
• The user message bubble and attachment chip text use the primary color and are legible.
• Linux: tao's draggable Wayland titlebar is preserved and follows the GNOME button-layout setting.

Localization

Italian was added to the in-app language switcher, and the Catalan locale code was corrected (ct → ca).

Engineering

• The MLX plugin and extension are built on macOS only, keeping tauri-plugin-mlx and its guest-js bindings out of Linux and Windows builds.
• The llama.cpp extension test suite now runs as part of the global yarn test.

Update your Jan or download the latest (opens in a new tab).

For the complete list of changes, see the GitHub release notes (opens in a new tab)