Developer Builds Full LLM System from Scratch Using Only Python and NumPy

A developer has released Draco AI, a fully open-source large language model system built entirely in pure Python and NumPy, with no dependency on frameworks like PyTorch or HuggingFace. The project covers the complete execution path, including both an Inference Engine and a Thinking Engine, and is accompanied by a pytest validation suite. Key technical features include ternary weight quantization that reduces memory usage by up to 80%, a three-tier hierarchical cache for efficient context retrieval, and a 12-stage cognitive pipeline with nine independent hallucination verifiers. The architecture enforces strict invariants such as one-way data flow, consistent positional encoding, and deterministic rollback in speculative decoding to prevent silent errors. The project is publicly available on GitHub, with an accompanying Discord community for contributors and users.