Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-09-21. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (open source, Apache-2.0) | 2026-09-21 |
| 2 | your GPU time | 2026-09-21 |
One-liner: MiniMind is an open-source project whose stated goal is to go from zero to a working ~64M-parameter language model on an ordinary personal GPU in roughly 2 hours of training and a few dollars of compute cost — and to show every step of the way.
Verified facts (as of 2026-09-21)
| Attribute | Value | Source |
|---|---|---|
| GitHub stars | 61,924 | repo page snapshot, 2026-09-21 |
| License | Apache-2.0 (LICENSE file read directly) | repo / raw LICENSE |
| Last commit | 2026-09-20 | commits feed |
| Headline claim | ~64M-parameter model trained from scratch in ~2 hours on a personal GPU | repo README |
| Pipeline covered | MoE, data cleaning, pretrain, SFT, LoRA, RLHF (DPO), RLAIF (PPO/GRPO/CISPO), tool use, agentic RL, distillation | repo README |
| Implementation | Core algorithms written from zero in native PyTorch, without third-party high-level training frameworks | repo README |
| Model family | MiniMind-V (vision), MiniMind-O (omni), MiniMind-dLM (diffusion LM), MiniMind-Linear | repo README links |
What it is actually for
MiniMind is a teaching repository, not a model you deploy. The README’s framing is deliberate: a 64M model is about 1/2700 the size of GPT-3, small enough that one person with one consumer GPU can run, break, and re-train the whole stack in an afternoon. Everything above the training loop — the data pipeline, the loss curves, the RL alignment stages — is exposed in readable, dependency-light code, which is exactly what most “how does LLM training work” questions require and what 70B-parameter research repos never show.
The catches
- The output model is a toy. A 64M-parameter model produces character-level, low-quality text. The value is the code path, not the checkpoint.
- README is Chinese-first with an English translation (README_en.md); docs drift between the two is always possible on a fast-moving repo.
- Hardware claims are the author’s — “2 hours on a personal GPU” is the project’s own benchmark framing, not an independently reproduced number, and cost depends on whatever GPU you actually have.
Who should use it
Students, self-taught engineers, and anyone who wants to understand pretraining → SFT → RLHF end-to-end by running it themselves instead of reading papers. Skip it if you need a deployable model; use it if you need a deployable mental model.
Star count, license text, activity, and feature claims verified against primary sources (GitHub repo page, raw LICENSE/README, commits feed) on 2026-09-21. This page follows an official-sources-only review: the training time and cost claims were not independently reproduced, so no hands-on assertions are made.