Ecosystem Vision

Description

INTELLECT-3 is a 106B parameter Mixture-of-Experts model trained by Prime Intellect using a two-stage recipe (supervised fine-tuning followed by large-scale reinforcement learning) on a 512 NVIDIA H200 GPU cluster over roughly two months. It was trained end-to-end with prime-rl, Prime Intellect's asynchronous, production-scale RL framework, using agentic RL environments built with the verifiers library and hosted on the community Environments Hub, spanning math, code, science, logic, deep research, and software engineering tasks. Training also relied on Prime Sandboxes, a high-throughput, secure code execution layer for agentic coding rollouts, and a custom compute orchestration stack (Ansible provisioning, Slurm/Cgroup v2, Lustre storage, DCGM/Prometheus observability) managing the H200 cluster. The full recipe — model weights, training frameworks, datasets, RL environments, and evaluations — is open-sourced, and the model can be used via a hosted chat interface (chat.primeintellect.ai) or Prime Intellect's Inference API, with Parasail and Nebius serving as inference providers.