I'm happy to announce the public release of NARC Genesys 5.0, the successor to NARC NEXT 4.1.
NARC is an individual project I started with the goal of building an efficient and highly optimized chess engine from scratch. It has also been my introduction to chess engine development.
Over the past three months, I've experimented with hundreds of ideas, with each change tested through SPRT matches. Most of these experiments were unsuccessful, and I wasted quite a lot of computing resources during the early stages because I didn't fully understand some of the concepts involved. However, the entire process has been a valuable learning experience. I also used tools such as claude code to implement and test ideas quickly so it definetely accelarated the development of the engine by a lot.
Genesys 5.0 is a complete rewrite of the original NARC engine/ NARC NEXT engine and can essentially be considered a full reimplementation.
Development and NNUE training
Development started with a hand-crafted evaluation (HCE), which was gradually replaced with NNUE evaluation.
The first NNUE was trained on only 1 million self-play positions and already outperformed the HCE version by approximately 100 Elo. I then repeated this process, gradually increasing the training dataset while improving the quality of the position labels using newer, stronger versions of the engine.
The final NNUE was trained on approximately 249 million filtered positions, combining the engine's own self-play data with positions from licensed sources, including selected human games from Lichess.
To avoid knowledge distillation from other authors' engines, all training positions were relabelled using my own engines, primarily NARC Genesys itself during development.
One of the main things I learned from these experiments is that the biggest strength improvements came from improving the quality and diversity of training data and optimizing the NNUE architecture.
Playing strength
Based on my own testing, NARC Genesys 5.0 has an estimated playing strength of approximately 3600 Elo on the CCRL Blitz scale and haven't tested it for long time controls . I would estimate its 40/15 rating will be anywhere between 3500 - 3550 .
What I find particularly interesting is that this level of performance was achieved with approximately 249 million training positions, which I curated in a span of 3 months.
The project was developed with limited computing resources, so efficient use of hardware and training data was an important part of development.
I don't think Genesys has reached its strength limit yet. Further improvements are currently constrained mainly by available computing resources and my own understanding of engine development.
Source code
GitHub: https://github.com/Madushan996/NARC
Download the newest release
https://github.com/Madushan996/NARC/releases/tag/v5.0
The source code is publicly available for anyone interested in examining the implementation.
This project has taught me a great deal about chess engine development, and I hope to continue improving it.
I hope testers find this engine interesting ,
Feedback, suggestions, and independent testing are very welcome.
Thanks,
Madushan