Skip to content

Releases: whit3rabbit/openkind

v0.4.0

Choose a tag to compare

@github-actions github-actions released this 14 Oct 19:29
Release v0.4.0

v0.3.0

Choose a tag to compare

@github-actions github-actions released this 14 Oct 03:22
Release v0.3.0

v0.2.0

Choose a tag to compare

@github-actions github-actions released this 04 Oct 16:04
OpenKind v0.2.0

OpenKind 0.1.0

Choose a tag to compare

@github-actions github-actions released this 04 Oct 00:31

OpenKind 0.1.0 is the first release of the Rust decision engine. It returns typed
Noul, Choice, and Score answers through Jev HTTP and gRPC interfaces,
with request-bound answer validation.

The release passed all 10 CI jobs, nine archive builds, nine clean-host smoke
tests, and three Homebrew installation checks. All 12 Rust crate archives were
verified from the clean release commit.

Included

  • The openkind CLI, openkindd daemon, and Rust client SDK.
  • Curated model installation, a local playground, backend diagnostics,
    pinned evaluation datasets, and a native benchmark harness.
  • Native CPU execution and optional MLX, CUDA, and ONNX backends.
    Linux ROCm provider selection is available in source builds.
  • TypeScript, Python, and Swift clients and local daemon wrappers from source.
    The Swift package follows the repository tag. npm and PyPI publication are
    outside this release.
  • API-key generation and validation before model loading.

Distribution

The release includes 12 Rust crates at version 0.1.0, standard and ONNX
archives for macOS, Linux, and Windows, a portable Linux musl archive, checksums,
and Homebrew packaging. Every Rust crate archive contains the MIT license.

Installation

brew install whit3rabbit/tap/openkind

For a Cargo installation, provide Rust 1.88 or newer, protoc, and native
C/C++ build tools:

cargo install --locked --version 0.1.0 openkind-cli openkind-server

See the model guide
for loadable profiles and backend requirements.

Qualification

Backend compilation, runtime discovery, model parity, task quality, and
throughput each require separate evidence. Offline and mock checks do not
establish model quality or accelerated parity. Checkpoint-backed and hardware
gates remain separate from this software release.