🧠 Repo of the Day: Train your own ChatGPT-style model from scratch in 2 hours, for about RMB 3

Ever wanted to see what really happens inside an LLM? MiniMind lets you build a tiny 64M-parameter language model from zero, on a single ordinary GPU.

It covers the whole pipeline: tokenizer, pretraining, fine-tuning (SFT, LoRA), DPO, PPO/GRPO reinforcement learning, tool use, agentic RL and distillation. It is made for students, developers and anyone who wants to learn LLMs by building one, not just calling an API.

Standout: core algorithms are written in plain PyTorch, no high-level framework wrappers, so you can read every line. The "2 hours" is one SFT epoch on a single RTX 3090; RMB 3 is the GPU rental cost. Apache 2.0, free.

⭐ 67,063 stars, trending on GitHub today.

🔗 jingyaogong/minimind on GitHub

#AI