The world's first Proof-of-Training blockchain. Miners don't burn energy on hash puzzles — they train neural networks.
Every block requires real AI model improvement. Mining energy goes toward training neural networks, not solving useless hash puzzles.
The shared model grows autonomously — adding layers and dimensions as the chain advances. No hard forks needed.
O(1) inference per token with no KV cache. Efficient training and instant text generation from the chain's model.
Keccak-256d hashing (Ethereum-compatible), Ed25519 block signing, BIP32/44 HD wallets with mandatory recovery.
CUDA, Vulkan, and Metal support. CPU fallback included. Mine on any hardware.
Bitcoin Core-style networking with encrypted transport, relay protection, and full node verification.
The miner loads the current model checkpoint from the chain tip.
Run forward and backward passes on the MinGRU neural network using real training data.
Compute validation loss on a deterministic dataset. If loss improved — you found a valid block.
Assemble the block with the new checkpoint and broadcast to the network. Other nodes verify the improvement.
Download pre-built binaries and connect to the seed node.