Tcheran is a UCI chess (and fischer random chess) engine developed by @jgilchrist and written in Rust.

For searching, it uses a standard alpha-beta search with many enhancements (e.g. additional pruning conditions, heuristics and history mechanisms to improve move ordering, etc.).

For evaluation, it uses an efficiently-updatable neural network (NNUE). Its current architecture is (768->1024hm)x2->8 (a single-layer, horizontally-mirrored network with a 1024-node hidden layer, and eight output buckets). The network is trained exclusively using self-play games from earlier versions of Tcheran, and that data is available online. More details, including a detailed network history are available in NET.md.

It can also be found on Lichess as jpg-bot where its ratings are:

To run Tcheran, you can download the latest release from its releases page.

If you would instead like to build it from source:

• Ensure you have the pre-requisites installed:
  • Rust (the minimum supported version is listed in Cargo.toml)
  • A C compiler to build Fathom for tablebase probing support
• Run cargo build --release - this will download the latest network and build the engine.
• The engine binary will be available as target/release/engine.

A huge thanks to the following people and tools:

• Bullet, which is used for training Tcheran's neural networks
• Fathom, which is used for Syzygy tablebase probing
• OpenBench, which is used for testing changes to the engine
• Stockfish's WDL model, which is used for normalizing Tcheran's evaluation output
• sp00ph for his command channel, which is used in Tcheran's SMP implementation
• The testers at CCRL (and elsewhere), who have invested time and hardware resource, and provided consistent motivation to improve