jingyaogong/minimind
64M-parameter LLM trained from scratch in 2 hours
- Chart
- Not on a chart right now
- Stars
- 66.6k
- Daily peak
- #14
- Days on chart
- 2
What it is
MiniMind is an open-source project for training a 64M-parameter language model from scratch using native PyTorch. It implements the full LLM pipeline—from data cleaning and pretraining to SFT, reinforcement learning, and tool use—without relying on high-level wrapper libraries.
Who it is for
It is built for developers and learners who want to understand and reproduce LLM internals and end-to-end training on a consumer GPU.
Why it is trending
The repository gained 188 stars today, reaching 65,693 total stars. It stands out by implementing core architectures and training algorithms from scratch rather than wrapping existing frameworks.
Main features
- Native PyTorch implementation of Dense and MoE architectures without high-level wrappers
- Full training pipeline covering Pretrain, SFT, LoRA, DPO, RLAIF (PPO/GRPO/CISPO), and Agentic RL
- OpenAI-compatible API server, Streamlit WebUI, and support for llama.cpp, vllm, and ollama
- Open-source pretraining and SFT datasets with custom tokenizer training scripts
Summarized on November 17, 2026 from the README and repository data. It can be wrong; check the source before relying on it. Report a wrong summary