As of September 2026, the UltraData series has grown to 10+ datasets spanning code, math, web data, and more, with 2.66M+ downloads to date.

Together, these releases make up a growing open-source data library for researchers and developers working on LLM training.

For MiniCPM5-2B, these datasets provide training signals for capabilities ranging from code and reasoning to early general-purpose agent behavior on the edge.

We don’t just open-source the weights. We open-source the data that helps make them possible.

🤗 Explore the data:

🔗 UltraData-Code
https://t.co/LQiZnqMyYo

🔗 UltraData-SFT-Agent-2609
https://t.co/FIV4B6keuf

🔗 UltraData-RL-2609
https://t.co/iW0qH7tHF3

🔗 UltraX-Preview
https://t.co/vdGHUZTdaF