Atomic Chat v1.1.79

🚀 New Features

MTP support on Windows via the upstream llama.cpp build. The Windows backend now uses the official ggml-org/llama.cpp build directly, bringing Multi-Token Prediction (MTP) — a speculative decoding technique that can boost throughput by 30–70% on supported models.
Connect to multiple MCP servers to bring your own tools and integrations into Atomic Chat. An in-app log viewer for MCP tool calls is included.
Tree view for project conversations in the sidebar.

Fixed GPU detection on Windows by bundling the missing CUDA 13 DLLs with the backend, which also improves CUDA performance and stability. (#14 — thanks @evil-doer and @lollo78)
Fixed Flash Attention crashing model startup: the on/off/auto value was being dropped when building llama-server arguments, so llama.cpp was consuming the next flag as the value.
Fixed the MSI installer to upgrade the previous version instead of installing alongside it. (#11 — thanks @abbaselmas)
Fixed blank window after auto-update on Windows 11 that required a manual restart.
Fixed a duplicate Llama.cpp provider stuck on "loading" after v1.1.76.
Fixed window control buttons not flush to the edge when the app is maximized on Windows 11.
The i and download buttons are now reachable via keyboard and screen readers.
Multiple stability and UI fixes.

Big thanks to @Vect0rM for code contributions, and to @evil-doer, @lollo78 (#14) and @abbaselmas (#11) for the detailed bug reports.