Prime Intellect

Prime Intellect is a San Francisco company building infrastructure for decentralized training and reinforcement learning, releasing open INTELLECT models and publishing its research agenda in the open. In this corpus it first appears declaring Recursive Language Models the paradigm of 2026, the inference scheme in which a model manages its own context from a Python REPL. It returns as an operator of the recursive loop rather than its subject, handing idle compute to Codex and Claude Code to beat the human NanoGPT speedrun baseline and shipping an open agent that rewrites itself mid-task past the human expert baseline.