Skip to content

v1.14.4

Choose a tag to compare

@mohitsoni48 mohitsoni48 released this 04 Oct 11:14
· 685 commits to main since this release
795c3a1

Added

  • Add your own engine by uploading its .zip. In the Add-engine flow, "Upload a .zip…" now sits beside
    "Choose folder…": pick a fork's release archive and TurboLLM searches it at any depth for llama-server,
    extracts the binary together with the runtime libraries it needs (.dll / .dylib / .so and versioned
    sonames, found anywhere in the archive, symlinked library chains included), and probes it — the same
    confirm-and-register step as a folder. The files land under TurboLLM's own engine storage as
    engines/build/<zip-name>/, laid out flat so the Windows loader and the library search path the engine
    already launches with resolve them — including on Termux/Android — and the engine's Delete button removes
    them, exactly like a self-built engine. Re-uploading a same-named zip updates that engine in place, and a
    failed or wrong-platform re-upload leaves the working install untouched. Official llama.cpp Windows zips
    work (a CUDA build without the separate cudart zip gets a clear warning); on Linux, macOS, and Termux the
    official release assets are .tar.gz, so there it covers fork zips. The packaged Android app declines with
    a clear message (its OS forbids running downloaded binaries). Corrupt, encrypted, or oversized archives
    fail cleanly and never leave a half-installed build behind.

Fixed

  • Copy buttons now work when TurboLLM is opened over your network or through a tunnel. Browsers only allow
    the clipboard on HTTPS or on the machine itself, so from http://192.168.x.x:6996 every Copy button (chat messages,
    code blocks, share links, the build commands, the API endpoint in onboarding, and the rest) silently did nothing.
    They now fall back to a selection-based copy that works on plain http and puts the text on the clipboard of the
    device you clicked on, not the computer running TurboLLM.