Prime Intellect announced prime‑rl version 0.10.0 together with verifiers 0.4.0. The update makes the new IPO (INTELLECT Policy Optimization) loss the default for asynchronous reinforcement learning, replacing the older DPPO‑based loss. It also includes performance tweaks that make training up to 1.15× faster on large models like Qwen3‑235B‑A22B, and adds support for new architectures such as DeepSeek V4 Flash and Qwen‑VL.
Why it matters
Researchers can train large language models a bit quicker and with a more stable RL loss, reducing the chance of training collapse.
