OpenShogiAI (OSAI)

日本語はこちら

OpenShogiAI is an independently implemented shogi (Japanese chess) engine and machine-learning research project, written in Rust with native and WebAssembly runtimes. It plays book-free shogi with a learned evaluation network of our own design (OSAVAL03) and ships the training and evaluation tooling used to produce the published models.

The browser application for playing and analysis is a separate repository: OpenShogiUI (OSUI). This repository (OpenShogiAI) owns the engine, the model format, training, and model distribution.

Project by TKY-27.

Features

• Full shogi rules: legal-move generation with checks, drops, promotions, repetition and stalemate handling, SFEN/USI/CSA support.
• Two runtime builds from one code base: a native USI command-line engine and a pure WebAssembly module used by OSUI.
• Pure-learned evaluation: the OSAVAL03 container carries a dual king-relative clipped-pair accumulator network (W256/W512) with direct cp and win-draw-loss heads. The pure-only runtime rejects any handcrafted, book or teacher fallback — play strength comes only from the trained weights.
• Book-free play everywhere: no opening books, fixed first moves, position tables or online engines, in development matches and in production alike.
• Native/Wasm parity: deterministic Wasm regeneration and cross-runtime evaluation checks are part of make check.
• Training pipeline: self-play generation, offline fixed-depth teacher labeling, dataset preparation with split/lineage control, and bounded optimizer runs with resumable state.

Models

Representative trained weights are distributed from the releases page (tag models-v1) as CC BY 4.0, including the newest generation OSAI R4 and earlier development generations. Model generations are listed newest to earliest; this is development order, not a strength ranking. No human-dan validation has been performed.

• What is in a release: weight files (OSAVAL03), the shared browser engine runtime, the runtime profile, per-model rights and source records, an Apple Silicon macOS USI binary, and checksums.
• Weights license and training-data attribution: docs/model/distribution.md.
• Format specification: docs/model/OSAVAL03_FORMAT.md.

Playing in the browser: OpenShogiUI.

Build and run

Requirements: stable Rust (>= 1.89), wasm32-unknown-unknown target, wasm-bindgen CLI 0.2.127, Python 3.12 with uv, GNU Make, and Node.js for the actual-Wasm checks.

git clone https://github.com/TKY-27/OpenShogiAI.git
cd OpenShogiAI
./scripts/bootstrap_macos.sh        # optional: verifies/installs local tools
uv sync --locked --group dev
rustup target add wasm32-unknown-unknown
cargo install wasm-bindgen-cli --version 0.2.127 --locked --root local/tooling/wasm-bindgen-0.2.127
make build

Run the default (handcrafted) engine:

cargo run --locked -p open-shogi-cli -- usi

Run a learned model in pure USI mode (weights from a release):

make pure-build
target/pure/release/open-shogi-cli usi \
  --model osai-r4.osaval03 \
  --model-sha256 9466a7e8cf11b7d165b325edd9a5a421bdbaa4bed550940afcf33c9faf3bfd0f \
  --model-format OSAVAL03 --profile pure_learned

All-in-one verification (format, lint, tests, native build, deterministic Wasm regeneration, license/boundary/provenance checks):

make check

The macOS binary in the release is an unsigned local build; see the release notes for the Gatekeeper note, or build from source as above.

Documentation

• Development guide — commands, storage, validation.
• USI / ShogiHome play guide — run the published OSAI R4 weights as a native USI opponent on Linux.
• Architecture and interface contracts.
• Rules profile — the shogi rules implemented.
• Model handling, OSAVAL03 format, distribution and weights license.
• Data rights and lineage, source audits.
• License scope and third-party notices.
• References — rules, algorithms and standards consulted, including the prior engines this independent implementation learned from (Apery, YaneuraOu lineage materials and others).

Contributing

Issues and pull requests are welcome; see .github/CONTRIBUTING.md. This is a solo-maintainer project: continuous updates, response deadlines and merging of every PR are not guaranteed, but quality discussions and contributions are actively considered. AI-assisted contributions are limited to agents with capabilities equivalent to or greater than Astra or Fable.

License

Project-owned source code is AGPL-3.0-only (LICENSE). Dependencies, datasets, third-party assets and model weights keep their own terms; see license scope and third-party notices. Published weights are CC BY 4.0 with the attribution recorded in docs/model/distribution.md.

Contact: X @ANAg2bGOD (DM).