fly-chess brains

Chess-playing neural networks whose wiring is the FlyWire adult Drosophila melanogaster connectome: 134,209 neurons, 2,700,513 neuron-to-neuron connections (34.2 M synapses), every synapse's sign fixed by the presynaptic neuron's neurotransmitter (acetylcholine excitatory; GABA and glutamate inhibitory). Training learns only synaptic strength magnitudes, per-neuron biases / leaks / gains, and the board-input and move-readout projections — no connection is added, removed or re-signed.

Code, training pipeline and the website that runs these brains in the browser: https://github.com/cesp99/fly-chess (MIT).

Model description

The network is a recurrent rate model on the connectome: h_t+1 = (1-a) h_t + a · f(W h_t + bias + input), W sparse with the connectome's pattern and signs, unrolled 8 or 16 timesteps per position. The board (20 planes × 64 squares, always from the side to move) enters through learned projections into sensory neurons; the move policy (4,168 logits) and value are linear/MLP read-outs of the descending and motor neurons. fly3 additionally lets the fly see the board: 5,543 real photoreceptors (R1–6, R7, R8, placed on their ommatidial columns; left eye files a–d, right eye e–h) each receive one square, and the signal travels through the real lamina → medulla → lobula → central brain wiring. Details: docs/SPEC.md and docs/RETINA.md in the code repository.

Files

file	step	stage	size	notes
checkpoints/fly1-imitation.pt	110,767	imitation	69 MB	random +191, material +70
checkpoints/fly2-imitation.pt	280,000	imitation	69 MB	random +338, material +70
checkpoints/fly3-imitation.pt	260,000	imitation	114 MB	random +338, material +241
checkpoints/fly3-selfplay.pt	264,000	selfplay	114 MB	random +636, material +241, previous-best +61
graph/full.npz	—	—	37 MB	the BrainGraph (CSR connectome, signs, retina mapping) every checkpoint needs
web/brain.json, web/brain.flyb, web/brain.flyb.gz	—	—	60 / 50 MB	browser blob of fly3-selfplay (SPEC §8 format; the website loads it from here)

Each checkpoints/<name>.json holds the exact training configuration and evaluation numbers of that checkpoint; manifest.json lists sizes and sha256 for every file.

Generations

run	recipe	held-out top-1 / top-3 (same 5,120 human positions)
fly1-imitation	8 timesteps, ReLU, board via 2,048 sensory/ascending neurons, Lichess 2014	34.1% / 58.4%
fly2-imitation	16 timesteps, saturating rates, Lichess 2014+2015, fp32	33.1% / 57.2%
fly3-imitation	fly2 + retina input, homeostatic gains, multi-timestep readout, neuromodulatory gating, central-brain readout, + 30 M Stockfish-evaluated positions	36.9% / 62.3%
fly3-selfplay	fly3-imitation + gated self-play (5 of 9 iterations promoted)	37.6% / 63.4%

Head-to-head with 100-simulation MCTS over 100 games on 50 paired openings: fly3-selfplay vs fly1 +50 =50 −0, vs fly2 +100 =0 −0; vs its own imitation checkpoint (200 sims, 20 games) 20–0; vs a 1-ply material-greedy bot +38 =2 −0 with search and +16 =4 −0 without. It is a beatable club-level-ish opponent without search and a real fight with it; it is not an engine.

Use

pip install git+https://github.com/cesp99/fly-chess
hf download cesp99/fly-chess --local-dir fly-chess-models

import torch, chess
from flychess.connectome.graph import BrainGraph
from flychess.model.config import BrainConfig
from flychess.model.flybrain import FlyBrain
from flychess.play.engine import FlyEngine

ck = torch.load("fly-chess-models/checkpoints/fly3-selfplay.pt", map_location="cpu", weights_only=False)
graph = BrainGraph.load("fly-chess-models/graph/full.npz")
model = FlyBrain.from_checkpoint(ck["model"], BrainConfig.from_dict(ck["brain_config"]), graph).eval()
engine = FlyEngine(model, graph, device="cpu")          # "cuda" if available
move, info = engine.choose_move(chess.Board(), difficulty="fly")   # larva | fly | superfly (MCTS)
print(move, info["policy_top"], info["value"])

Or with the CLI from the code repository: fly play --ckpt fly-chess-models/checkpoints/fly3-selfplay.pt --gui.

Training data

• Lichess monthly databases 2014-01 … 2015-12 (CC0): rated standard games, both players ≥ 1800 Elo, not bullet; 229 M positions (policy target = the human move, value target = the game result).
• Lichess evaluation database: 30 M positions with Stockfish evaluations (policy target = engine best move, value target = win probability from the centipawn score); used for fly3.
• Self-play games of the network itself (gated, with human/engine positions rehearsed in every batch).

Limitations

A fixed random-for-chess topology with sign constraints is a poor chess architecture: the brains play plausible openings and positional chess and still miss tactics; value estimates are weak without search. The rate model is a caricature of real neural dynamics (no spikes, gap junctions, dendrites or real neuromodulation). Input/output neuron choices are modelling decisions, not biology.

Licence and attribution

The FlyWire Codex data release the weights derive from is CC BY-NC 4.0, so the weights are published under CC BY-NC 4.0 (LICENSE): non-commercial use only, with attribution. See ATTRIBUTION.md for the full list of sources and citations. Training code: MIT.

Citation

@misc{flychess2026,
  title  = {fly-chess: a chess engine wired like the FlyWire fruit-fly connectome},
  author = {Esposito, Carlo},
  year   = {2026},
  url    = {https://github.com/cesp99/fly-chess}
}

Please also cite the connectome papers listed in ATTRIBUTION.md.