prime-rl 0.10.0: faster training, our new RL loss IPO, and one unified eval stack

Prime Intellect released prime-rl 0.10.0 and verifiers 0.4.0, introducing IPO (INTELLECT Policy Optimization) as the default asynchronous reinforcement-learning loss, alongside performance improvements, expanded model support, and a unified evaluation workflow. IPO replaces the previous DPPO-based default by masking tokens according to absolute probability differences between training and inference rather than probability ratios, which the developers report improves stability in extreme off-policy agentic SWE RL tests. The release adds maximum-throughput training recipes and optimized MoE, sparse-attention, and custom kernel implementations that improve SFT performance for models including Qwen3-235B-A22B, GLM-5.3, and DeepSeek V4 Flash. It also supports new architectures such as Qwen3.8 Flash Next and Nemotron-H, introduces multimodal RL support for Qwen-VL and Kimi K2.5, and adds RL features including exact sampling replay, constant batch backfilling, and default MoE-router freezing. Evaluation tooling has moved into prime-rl through a new eval CLI supporting adaptive concurrency, resumable runs, shared task selection, and distinct handling of timed-out rollouts, while verifiers 0.4.0 completes the transition to its v1 CLI stack.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.