MiniMind enables anyone to train a 26M-parameter GPT-style language model from scratch in just 2 hours on a single GPU, dramatically lowering the barrier to LLM education and experimentation. The project covers the full training pipeline including pretraining, SFT, RLHF with PPO/GRPO/SPO, MoE architectures, and native agentic RL with multi-turn tool-use. The latest minimind-3 release aligns its architecture with Qwen3/Qwen3-MoE featuring Dense ~64M and MoE ~198M/A64M models with updated tokenizer supporting tool call and thinking tokens.