DigiNews

Summary

MiniMind is an open-source project aimed at training a tiny 64M-parameter LLM from scratch with minimal cost. The README details the architecture, training data (pretrain, SFT, RL), and multiple RL methods (DPO, PPO, GRPO, CISPO, Agent RL), plus tooling and deployment options for local inference and OpenAI-compatible API. It emphasizes self-hosted, low-cost experimentation for individual developers and SMBs.