A YouTuber named Sammyuri has replicated a ChatGPT-style language model inside Minecraft, a build dubbed CraftGPT that showcases what is possible when redstone and patience meet machine learning. However, the result is minimal and very slow.
CraftGPT’s neural network is relatively small by modern standards, with approximately 5 million parameters, and the entire build occupies 439 million blocks. To keep that enormous structure visible and running, the creator used the Distant Horizons mod, while claiming the model logic itself is implemented using vanilla Minecraft redstone mechanics.
sammyuri trained the model in Python using the TinyChat dataset on HuggingFace. That dataset, combined with the model’s size, means CraftGPT’s vocabulary is just 1,920 tokens, and it has a context window of only 64 tokens, so its output is quite limited.
On the project’s GitHub page, the creator warns readers not to expect polished conversation. As they write, “you shouldn’t have high expectations. The model is very prone to going off topic, producing responses that are not grammatically correct, or simply outputting garbage. The model also has a very small context window of only 64 tokens. The conversations in the showcase video show the model at its best, not necessarily at its average performance.”
Performance is another major limitation. Even with an accelerated tick rate provided by Minecraft High Performance Redstone, sammyuri says CraftGPT can still take hours to generate a response. Without that multithreaded server setup, the creator estimates response times could stretch to decades.
Share your thoughts in the comments, and find us on X and Bluesky.
