Skip to content

Releases: Demorolled/VACA

Veronica 1.0.0 — Linux x64 · Windows x64

Choose a tag to compare

@Demorolled Demorolled released this 09 Oct 14:54

Veronica 1.0.0 — Linux x64 · Windows x64

First public build of Veronica (VACA), the AI code architect: describe an app in
plain English and it builds it.

This is a time-limited demo. It runs normally until 9 December 2026,
then refuses to launch. The date is fixed rather than counted per machine, so
every copy stops on the same day. Nothing is deleted when it expires — see
"Demo builds" in the README for exactly what it does and does not do.

Downloads

File Use
veronica_1.0.0_amd64.deb Ubuntu/Debian system install. sudo apt install ./veronica_1.0.0_amd64.deb
Veronica-1.0.0.AppImage Portable, no install. Needs libfuse2.
Veronica-1.0.0-linux-x64.tar.gz Self-contained runtime directory. Extract anywhere and run linux-unpacked/veronica.
Veronica-Setup-1.0.0.exe Windows x64 installer. Per-user NSIS, no admin prompt, creates shortcuts. Unsigned.
SHA256SUMS Checksums for the four files above.

You also need the model — it is not in these files

The weights are ~9 GB and are deliberately not bundled. Install them once per
machine, separately:

./scripts/install-model.sh                    # public qwen2.5-coder:14b
./scripts/install-model.sh --tuned --gguf-url <URL>   # the VACA-tuned Qwen2.5 14B

See the README section “The model — downloaded separately”. Briefly:

  • The app ships pinned to qwen2.5-coder:14b, served over the
    OpenAI-compatible API (Ollama by default, http://127.0.0.1:11434/v1).

  • This build was developed against a tuned Qwen2.5-Coder-14B-Instruct-Uncensored
    Q4_K_M
    , created in Ollama as vaca-r20-q4. It is hosted on Hugging Face, not
    attached here — see The tuned model in the README:
    https://huggingface.co/StevenWoods/veronica-r20-q4-GGUF

    ollama run hf.co/StevenWoods/veronica-r20-q4-GGUF

    That repository carries the template and params files, so the one command
    above is the whole install. ./scripts/install-model.sh --tuned does the same
    thing through a local Modelfile.

  • Without a reachable model the UI loads and the app runs, but generation does
    nothing: /health reports llm.up:false.

Requirements

  • Linux: Ubuntu/Debian x86-64. GTK3, NSS, libsecret and the usual Chromium
    runtime libraries (the .deb declares them). Windows: 10/11 x64.
  • The Windows installer was cross-built with wine on Linux and its packaged
    payload was verified, but it has not been installed and launched on a real
    Windows machine. Expect to shake out first-run issues.
  • No system Node — the bundled backend runs on Electron's own Node.
  • A GPU is not required, but a 14B Q4_K_M wants roughly 9–10 GB of VRAM to be
    usable. It will fall back to CPU, slowly.

Known issues

  • On Ubuntu 24.04+ the app may abort at startup with
    The SUID sandbox helper binary was found, but is not configured correctly.
    Run with --no-sandbox, or restore the setuid bit / install the AppArmor
    profile the build ships with — all three fixes are in the README's
    Troubleshooting section. This affects the .deb's menu entry too.
  • Unsigned binaries; some systems will warn.
  • chrome-sandbox in the tarball is not setuid, so linux-unpacked/veronica
    needs --no-sandbox (or the same fix).

License

Proprietary — all rights reserved. Source is published for reference; that is
not a license to reuse it. Electron, Chromium and the vendored packages keep
their own licenses, and the model carries its own.


Verifying what you downloaded

SHA256SUMS covers the four files above. From the directory you downloaded them
into:

sha256sum -c SHA256SUMS

The model is a separate download and has its own checksum on Hugging Face.