Hyperion-Chess

Chess engine for AI Chessathon 2026.

Overview

Hyperion-Chess is a competition-grade Python chess agent combining:

  • Iterative deepening alpha-beta with quiescence search
  • Numba-jitted PST evaluation (Phase 1)
  • Micro-NNUE int8 weights (Phase 3, trained locally on SF-labelled FENs)
  • 4-man Syzygy tablebases for perfect endgames within the 50 MB zip cap
  • Dynamic time management for 120s + 0.5s control

Source code

GitHub: shahaman098/hyperion-chess

Weights (coming soon)

File Description
nnue.npz Quantized Micro-NNUE (768โ†’128ร—2โ†’32โ†’1)
Syzygy 4-man Endgame tablebase (too large for HF; shipped in competition zip)

Usage

This is a competition submission, not a pip package. The platform imports agent.get_move(fen, time_left_ms).

import agent
move = agent.get_move("rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1", 120_000)
# โ†’ "e2e4"

Training

Training scripts live in the GitHub repo under training/. Positions may be labelled with Stockfish at train time (allowed by competition rules); only self-trained weights ship in the zip.

License

Apache 2.0

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support