Nawal
A 100% sovereign AI system with federated learning, genome evolution, and hybrid teacher-student architecture. Nawal ensures that AI capabilities remain under national control while contributing useful work to the network.
Key Features
NawalTransformer
A 100% sovereign language model built from scratch in 117M, 350M, and 1.3B parameter variants — no external model weights or dependencies.
Genome Evolution
DNA encoding, fitness evaluation, and genetic operators for evolving model architectures — 3,185 lines of evolution logic.
Federated Learning
Flower-based federated learning with a custom FedAvg aggregator, differential privacy, and Byzantine-robust aggregation. Includes robust workload simulation and metrics collection tools. Data never leaves local nodes.
Hybrid Teacher-Student
DeepSeek serves as an external teacher while NawalTransformer acts as a sovereign student, learning without data exposure.
PoUW Integration
AI training workloads are designed to contribute to Proof of Useful Work consensus so validators earn rewards for useful computation (on-chain integration prepared; not yet live on the Ceiba testnet).
FastAPI REST API
Full HTTP API for inference, training management, federated coordination, and model health monitoring (port 8080, fronted by /api/nawal/ on the Ceiba testnet).