Release2026-08-07
Prime Intellect extended its open-source reinforcement learning tools so several models can be trained together while interacting, rather than one model at a time. Four ready-made setups ship with it, including one model grading another's work, a model playing a game against itself, and a simulated user talking to an assistant. The work is aimed at machine learning researchers and requires the new releases of the verifiers and prime-rl libraries.
What changed
The stack supported training and evaluating one agent at a time on a fixed task set.
What it unlocks
Training and evaluating several models that interact with each other in one run, choosing which roles learn and scoring each role separately.
- verifiers 0.3.0
- prime-rl 0.8.0
- four multi-agent environments implemented
- learnability peaks at a 50% solve rate
What you need to act on it
- open-source verifiers 0.3.0 and prime-rl 0.8.0
- reinforcement learning training infrastructure
- primeintellect.ai2026-08-07