MikeTrendsTrends right now

⬢github Python · 66.3K ★ +636 since we first saw it · pushed 1 h ago · Apache-2.0

jingyaogong/minimind

🧠 Train a 64M-parameter LLM from scratch in just 2h!

MiniMind is an educational open-source project that lets you train a tiny (~64M-parameter) LLM entirely from scratch in about 2 hours on a single consumer GPU, for a few dollars of cloud cost. It includes the model architecture (Dense and MoE), tokenizer training, full-pipeline datasets, and complete code for pretraining, SFT, LoRA, DPO, PPO/GRPO, distillation, and more — all implemented natively in PyTorch without high-level framework abstractions.

Why now: It has become a hugely popular learning resource (65k+ stars) for its low-cost, fully reproducible from-scratch LLM training pipeline, and it keeps expanding with new variants like vision and multimodal models.

Who it is for: Developers and students who want to understand how LLMs work internally by building and training one themselves, rather than just fine-tuning existing models.

llmpytorcheducationtraining-from-scratchmoerlhf

artificial-intelligencelarge-language-model

Open on GitHub →

Stars over our 82 snapshots: 65.6K to 66.3K, since 20 h ago.

Where people talked about it

API: https://socialmediatrends-api.osmike.com/v1/repos/jingyaogong/minimind