Installation

Spark runs on Linux, macOS and Windows, on x64 (AVX2 or AVX-512) and ARM64 CPUs. Models run on the CPU unless an NVIDIA GPU with CUDA is present. Releases are self-contained, so no .NET installation is needed.

Install script

curl -fsSL https://raw.githubusercontent.com/curiosity-ai/spark/refs/heads/main/install.sh | sh

The script downloads the latest release for your platform, checks its SHA-256 and installs the program in ~/.curiosity/spark.

Linux and macOS: a launcher goes in ~/.curiosity/bin (shared with curio and the Orchestrator), and your shell profile gets that folder on PATH.
Windows: the program folder is added to your user PATH.

Open a new shell afterwards so that spark is on the path.

Install settings

All settings are optional environment variables.

Variable	Meaning
`SPARK_VERSION`	Install a specific release, such as `v26.10.6452`, instead of the latest.
`SPARK_INSTALL`	Install under another folder instead of `~/.curiosity`.
`SPARK_NO_MODIFY_PATH=1`	Leave `PATH` and your shell profile alone.
`GITHUB_TOKEN`	A GitHub token, for when you hit the API's anonymous rate limit. `GH_TOKEN` works too.

curl -fsSL https://raw.githubusercontent.com/curiosity-ai/spark/main/install.sh | SPARK_VERSION=v26.10.6452 sh

Upgrade

Run the install script again. It replaces the program folder. Your data and models are stored elsewhere (see SPARK_STORAGE), so they are kept. Restart Spark afterwards. On Windows, stop it first, because Windows locks the folder of a running program.

Manual download

Every release has an archive per platform: linux-x64, linux-arm64, osx-x64, osx-arm64 and win-x64. Check yours against the release's SHA256SUMS, then extract it and run it:

tar -xzf spark-<version>-linux-x64.tar.gz
cd spark-<version>-linux-x64
SPARK_ADMIN_PASSWORD='choose-a-password' ./spark

Keep the folder together. It holds the server (spark), spark-runner, the CUDA kernels, the native libraries and the web console (wwwroot/). The server finds the runner beside it, or at SPARK_RUNNER_PATH.

NVIDIA GPUs

CUDA on Linux needs the NVIDIA driver and cuBLAS 12 (libcublas.so.12, from the CUDA toolkit or the libcublas-12-* package). On Windows it needs the CUDA toolkit's cublas64_12.dll on the PATH. The install script warns when an NVIDIA GPU is present and cuBLAS 12 is not. The Docker image already carries cuBLAS. Without a CUDA device, Spark runs every model on the CPU. The device settings are in Runners and placement.

Check the install

SPARK_ADMIN_PASSWORD='choose-a-password' spark

Startup fails when SPARK_ADMIN_PASSWORD is missing or shorter than 8 characters. Open http://localhost:8080 and sign in as admin. The overview shows the version, the number of CPUs, whether a CUDA device was found and how much memory is free for models.

Next: Quick start