Neural Architecture

Edge-Native Chess Models

Lightweight Transformer neural network specialized for Chinese Chess. Millisecond-latency edge inference with zero backend compute overhead.

Active Live Version9.74M Params

Xiangqi Transformer V2.10

Equipped with 32-ply action history self-attention and batched PVS search. Formally verified past Pikafish UCI_Elo 2000 baseline.

Parameters9.74M (Edge-Native)
Action History32-ply Attention
Strength RatingPassed Pikafish 2000 Gate
Latency~8.5ms (Client Wasm)
Coming SoonNext Gen

Next-Gen Model in Training

A next-generation neural architecture with deeper tactical search and reinforcement self-play is currently in training. Stay tuned.

StatusTraining in Progress
ProgressComing Soon
PlayableCurrent Model Live