INTELLECT-3 is a frontier-class 100B+ Mixture-of-Experts language model developed by Prime Intellect and trained end-to-end using their large-scale asynchronous RL framework, PRIME-RL. Built on the GLM-4.5-Air base model, INTELLECT-3 combines supervised fine-tuning with long-horizon reinforcement learning across hundreds of verifier-backed environments spanning math, code, science, logic, and agentic tasks.

The model was trained on a high-performance cluster of 512 NVIDIA H200 GPUs across 64 nodes, supported by Prime Intellect’s Sandboxes execution engine, deterministic compute orchestration, and Lustre-backed distributed storage. The result is a model that surpasses many larger systems in reasoning benchmarks while remaining fully open-source.

Prime Intellect released not only the model weights but also the full training recipe: PRIME-RL, Verifiers, the Environments Hub, datasets, and evaluation suites. INTELLECT-3 is positioned as a foundation for organizations seeking to post-train or customize their own frontier-grade models without relying on proprietary AI labs.

A New Challenger: INTELLECT-3's 100B Parameters Punch Above Their Weight