November 2025 Summaries

INTELLECT-3, a 100B+ parameter Mixture-of-Experts model, sets a new standard for its size in benchmarks across math, code, science, and reasoning, surpassing many larger models. Developed using an open-source reinforcement learning (RL) stack by Prime Intellect, it demonstrates the potential for any company to engage in AI development. Built on the GLM 4.5 Air base model, INTELLECT-3 utilizes PRIME-RL, an asynchronous RL framework, and various infrastructures like Prime Sandboxes and the Environments Hub, allowing for efficient, scalable training across 512 NVIDIA H200 GPUs. The model's training involved supervised fine-tuning and large-scale RL, leveraging a diverse set of publicly available environments to enhance its reasoning and agentic capabilities. With resources and frameworks like prime-rl and verifiers open-sourced, Prime Intellect aims to democratize AI by enabling broader participation and innovation in the field, moving towards a future where AI development is accessible to all.