INTELLECT-3 is a 106-billion-parameter Mixture-of-Experts language model with roughly 12 billion parameters active per inference, designed to balance broad capability with efficient deployment. It was developed by the Prime Intellect team and released alongside its complete training infrastructure, including open-source reinforcement learning frameworks, the full training recipe, and a collection of training and evaluation environments. The model is built on top of the GLM-4.5-Air-Base architecture and was refined using both supervised fine-tuning and large-scale reinforcement learning, with RL training scaled across up to 512 H200 GPUs to push performance in math, code, science, and general reasoning.

According to the technical report, INTELLECT-3 achieves state-of-the-art results among models of its size across a range of math, coding, science, and reasoning benchmarks, outperforming several larger frontier systems. A central contribution is the prime-rl framework, an open asynchronous reinforcement learning stack that scales from a single node to thousands of GPUs and supports multi-turn agentic interactions with tool use. The model supports function calling and tool integration for practical deployment and is listed on third-party routing and comparison platforms, indicating availability for real-world applications in reasoning-heavy and code-oriented workflows.