Prime Intellect

Prime Intellect builds decentralized training infrastructure, open training environments and research agents. In this corpus it is the lab that gave Codex and Claude Code some 14,000 H200 hours of idle compute and let them beat the human baseline on the NanoGPT speedrun optimizer track, and five days later shipped the 4,504-task environment that tripled small-model tool-use scores through self-play. It is the same company as prime-intellect, the older id under which the corpus records its first appearance, the January Recursive Language Model; the two entries are one organization, pending a merge onto the older id.