Train your own LLM for 43 cents and 2 hours (minimind setup)

This is the setup for minimind, the project from the reel. Everything below is from its own README, read on 6 September 2026.

What it actually is: a 64M parameter language model you train yourself, from nothing, on one consumer GPU. Pretraining, then instruction tuning, then a model you can chat with. The whole point of the project is that the numbers are small enough to run on a card you might already own.

The number, and exactly what it means

The README's own cost table, one RTX 3090 rented at about 1.3 yuan an hour, 7 yuan to the dollar:

Stage	Time	Cost
pretrain on pretrain_t2t_mini	about 1.21 h	about 1.57 yuan
SFT on sft_t2t_mini	about 1.10 h	about 1.43 yuan
one epoch of both	about 2.31 h	about 3.0 yuan, roughly 43 US cents

Three things the README says out loud, so this page says them too. That is one epoch of the MiniMind Zero variant, on the two small datasets. It is an estimate, and the author calls it "a rough reference for training cost". And the author's own rig is eight 3090s, so the single card figures are scaled, not clocked.

Step 0, clone and install

git clone --depth 1 https://github.com/jingyaogong/minimind
cd minimind && pip install -r requirements.txt

Check the GPU is visible before anything else:

python -c "import torch; print(torch.cuda.is_available())"

The author's environment for reference: Ubuntu 20.04, CUDA 12.2, Python 3.10.

Step 1, the two data files you actually need

Only two files reproduce the Zero model. Download them from ModelScope or Hugging Face and put them in ./dataset/:

• pretrain_t2t_mini.jsonl, 1.2GB
• sft_t2t_mini.jsonl, 1.6GB

ModelScope: modelscope.cn/datasets/gongjy/minimind_dataset

Hugging Face: huggingface.co/datasets/jingyaogong/minimind_dataset

You do not need the rest of the dataset for this.

Step 2, train it

cd trainer && python train_pretrain.py
cd trainer && python train_full_sft.py

Both save checkpoints to ./checkpoints/. If a run dies, add --from_resume 1 and it picks up where it stopped.

Step 3, talk to it

Weights land in ./out/. Then:

python eval_llm.py --weight full_sft

Skip the training, just run the model

If you want the chat model without the two hours:

git clone https://huggingface.co/jingyaogong/minimind-3
python eval_llm.py --load_from ./minimind-3

What to expect, honestly

It is tiny. 64M parameters is about one 2,700th of GPT-3 by the README's own comparison. This is a model that teaches you how training works, and it will chat, and it will not replace anything you pay for.

The 2 hours is one epoch on the mini data. More data, more epochs and the MoE variant all cost more; the README's table puts the 198M MoE at about 3.2 hours and 4.2 yuan for the same pass.

The cost is a rental price, and rental prices move. Measure your own wall time. If you own the card, the cost is electricity.

Multi GPU shortens it. torchrun --nproc_per_node N on any of the training scripts, and the README says an 8x H100 setup gets the whole thing down to minutes.

The source

The project: github.com/jingyaogong/minimind

58,854 stars, Apache 2.0, first commit July 2024. Every claim on this page is from that README, read on 6 September 2026. If it has changed since, trust the repo over this page.

---

What did your wall time say?

Train one on whatever card you have, then tell me in The Rabbit Hole what it actually took, because the only figure that matters here is the one clocked on your machine and not estimated on somebody else's.

It is the room where I keep the parts that don't fit on a page like this: the full run of numbers, what broke the first time, and the ones that looked good and then didn't work at all.

Two catches, up front, because that's the rule here. It's private, so you have to be inside to read anything. And it's new, so you'd be early rather than joining something busy.

Free: skool.com/the-rabbit-hole-3206/about