kiwi

kiwi is a tiny k-like array language implementation that can lower to mlx and webgpu.

Get Kiwi

Download CLI binaries for macOS and Linux from GitHub Releases.

Apple apps for iOS, iPadOS, macOS, and watchOS are available on the App Store.

Try Kiwi

Try Kiwi in the browser at kiwilang.com/repl.

With the CLI installed:

kiwi
kiwi path/to/file.k

Running kiwi without a file starts the local REPL.

Jupyter

The Jupyter kernel lives in python/jupyter/. It uses the shared libkiwi_bridge session API and can display Vega-Lite output emitted as JSON through Kiwi's `j@ encoder.

In a fresh public checkout, bootstrap MLX first with KIWI_MLX_BACKEND=cpu scripts/bootstrap_deps.sh unless .deps/mlx and .deps/mlx-c already exist.

cd python/jupyter
uv sync --managed-python --python 3.14.2 --group dev
uv run --managed-python python -m kiwi_jupyter_kernel.install --user

Build From Source

Requirements

• Zig on PATH or ZIG_BIN set explicitly
• CMake for MLX dependency builds
• a supported host toolchain for Zig and MLX

Standalone Host CLI

Build the standalone host CLI:

scripts/bootstrap_duckdb.sh
zig build -Dpublic-cli=true -Druntime-backend=host -Dstrip-instrumentation=true -Doptimize=ReleaseFast

MLX Backend

Bootstrap MLX and build with the native MLX backend:

KIWI_MLX_BACKEND=cpu scripts/bootstrap_deps.sh
zig build -Druntime-backend=mlx

scripts/bootstrap_deps.sh now also bootstraps a repo-local DuckDB into .deps/duckdb. Host-matching builds prefer that install automatically so DuckDB-backed CSV, Parquet, and HTTP(S) scans work without runtime extension installation.

Dependencies

Pinned dependency revisions live in deps.lock.toml.

See THIRD_PARTY_NOTICES.md for dependency notices.