knaif 1.1.0

The first knaif release with downloadable binaries. 1.0.0 and 1.0.1 published the knaif wheel to PyPI only — no GitHub Release has ever existed, and the Windows artifacts built alongside them were never uploaded. Every artifact here is therefore unproven in the field by definition.

Planning is unchanged from 1.0.1; one expansion fix is not. No prompt, validation or model change — the same utterance yields the same plan. The single exception: converting to webm used to emit a stream-copy that ffmpeg rejects, leaving a truncated file behind; it now re-encodes to VP9 + Opus when the source codec cannot live in the target container. What makes this a minor release is the new knaif backend subcommand; everything else is packaging correctness and artifacts that should always have shipped.

Which download

Platform	File
Windows	knaif-1.1.0-windows-x64-setup.exe (per-user installer, no admin) or knaif-1.1.0-windows-x64.zip (portable)
Linux	knaif-1.1.0-linux-x86_64.AppImage (single file) or knaif-1.1.0-linux-x64.tar.gz (portable tree)

One artifact per OS, carrying both the CPU and Vulkan backends; the right one is selected at runtime.

Supported floor

Platform	Floor
Windows	x64, Windows 10 or later
Linux	x64, glibc 2.34+ and libstdc++ with GLIBCXX_3.4.30 / CXXABI_1.3.13 — Ubuntu 22.04+, Debian 12+, Fedora 36+, Mint 21+
Not supported	RHEL / Rocky / Alma 9 — glibc is new enough, libstdc++ is one version short

libstdc++, not glibc, is the binding constraint on Linux. Both directions are tested: the artifact must run at the floor and fail below it.

NVIDIA users — read this

knaif backend install cuda fetches an opt-in CUDA backend (~668 MB on Windows, ~731 MB on Linux; needs an R580+ driver). It takes effect on the next run.

On Blackwell (RTX 50xx) this is not optional. The bundled Vulkan backend generates at roughly CPU speed there — ~5.7 tok/s against the CPU's ~5.9 on knaif's own workload. That is not a slower option, it is a product that reads as broken. On older NVIDIA cards CUDA is faster and genuinely optional, and knaif says so in those terms rather than the alarming ones.

knaif offers it on first run when it finds an eligible GPU. A driver below the CUDA 13 floor gets an update hint and no offer, rather than a download that then fails to load. KNAIF_NO_CUDA_NUDGE silences the offer.

The backend is ABI-coupled to the binary that loads it, so an upgraded knaif refuses a payload left behind by an earlier release rather than loading it.

Verify your download

sha256sum -c SHA256SUMS --ignore-missing          # Linux
(Get-FileHash -Algorithm SHA256 <file>).Hash      # Windows — compare to SHA256SUMS

Windows SmartScreen. knaif ships unsigned, so Windows shows "Windows protected your PC" — More info → Run anyway. Verify the checksum first; that, not the absence of a warning, is what proves the download is intact.

First run

No --model needed. knaif offers to download the recommended GGUF (~2.5 GB) into ~/.knaif/models the first time you run it. Upgrading knaif re-downloads nothing while the recommendation is unchanged.

External tools are not bundled — ffmpeg, LibreOffice, Ghostscript and Tesseract install separately. knaif skills deps reports what each skill needs.

AppImage

chmod +x knaif-1.1.0-linux-x86_64.AppImage
./knaif-1.1.0-linux-x86_64.AppImage

Needs FUSE2 (libfuse2t64 on Ubuntu 24.04). Without it, run with --appimage-extract-and-run.

Unattended install (Windows) — pass /TYPE=full

The interactive installer defaults to the full type, but a /VERYSILENT install without /TYPE or /COMPONENTS reuses whatever component selection a previous install remembered — which can silently install only some skills. For deterministic scripted or CI installs, always pass the type:

knaif-1.1.0-windows-x64-setup.exe /VERYSILENT /SUPPRESSMSGBOXES /TYPE=full /DIR="<path>"

Add /TASKS="" to skip the PATH entry, winget dependencies, and the model download.

Uninstalling, and what happens to your data

~/.knaif — the model store plus the opt-in backends/ payload — lives outside the install directory, so it survives an upgrade.

• Windows: the uninstaller asks whether to delete ~/.knaif and deletes by default. Answer No only if you plan to reinstall and want to keep the ~2.5 GB model. Upgrades never prompt (Inno installs over an existing install). A /SILENT uninstall deletes without asking.
• Linux: the tarball and AppImage have no uninstaller. Delete the unpacked folder (or the .AppImage), and rm -rf ~/.knaif to reclaim the model.

Known limitations

No macOS artifact. Windows binaries are unsigned. The io skill is stale. External tools are not bundled (see First run above).

Full detail in CHANGELOG.md.

CUDA 13.3 redistributables

NVIDIA CUDA 13.3 redistributable runtime libraries for knaif's opt-in CUDA backend (knaif backend install cuda).

Keyed to the CUDA toolkit, not to a knaif version: these bytes are shared across knaif releases and this tag is never deleted, because installed backend manifests point at it permanently. The ABI-coupled ggml-cuda library rides the product release tag instead.

Redistributed under the NVIDIA CUDA EULA, whose text is included here and is also installed alongside the libraries.