Today, we're launching verifiers v1 (3 min read)

Prime Intellect has launched verifiers v1, a system optimized for agentic RL (reinforcement learning) and evals in modern environments. This system is designed to decompose environments into taskset, harness, and runtime, enabling complex coding and computer-use tasks to run at scale.

Key features include:

• Flexible configuration: You can reuse tasksets from existing frameworks like Harbor, or run evals and training across various harnesses such as Codex and mini-SWE-agent.
• Diverse runtime support: It can run in various environments including subprocess, Docker, and sandboxes.
• Efficient data management: By using a message DAG (Directed Acyclic Graph) structure, trace size has been reduced from the existing $O(n^2)$ to $O(n)$, enabling training on long-horizon agentic rollouts.

Additionally, verifiers v1 connects directly to prime-rl for use in training. In practice, applying the GLM-4.5-Air model to ScaleSWE tasks demonstrated high efficiency, completing 1,000 steps in just 2 days using only 6 H200 nodes.