Prime Intellect Unveils 106 Billion Parameter INTELLECT-3 AI Model, Claims State-of-the-Art Performance

Prime Intellect has announced the release of INTELLECT-3, a new 100B+ Mixture-of-Experts (MoE) model, marking a significant advancement in the field of large-scale artificial intelligence. The company states that the model achieves state-of-the-art performance for its size across critical domains including math, code, and reasoning, leveraging its proprietary end-to-end reinforcement learning (RL) stack. This development underscores Prime Intellect's ongoing mission to democratize AI development.

Prime Intellect, founded in 2023 by Vincent Weisser and Johannes Hagemann, operates as a decentralized AI platform. The company has successfully raised over $20 million in funding from prominent investors such as Founders Fund and CoinFund, focusing on enabling collaborative and open AI model training. Their platform aims to provide the infrastructure and tools necessary for broader participation in AI innovation.

INTELLECT-3 is a 106 billion parameter MoE model, which was post-trained from the GLM-4.5-Air base model through a combination of supervised fine-tuning (SFT) and extensive large-scale reinforcement learning. The training process utilized Prime Intellect's specialized PRIME-RL framework, Verifiers, Environments Hub, and Prime Sandboxes. This intensive training was conducted on a powerful infrastructure comprising 512 NVIDIA H200 GPUs distributed across 64 interconnected nodes.

According to Prime Intellect, INTELLECT-3 "achieves state-of-the-art performance for its size across math, code, science and reasoning benchmarks, outperforming many larger frontier models." Performance benchmarks shared on Hugging Face indicate strong results, with INTELLECT-3 scoring 98.1% on MATH-500 and 90.8% on AIME24, showcasing its capabilities in complex problem-solving.

The company emphasized its commitment to open-source AI, stating that INTELLECT-3 was "Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more." Prime Intellect has open-sourced the model weights, training frameworks, datasets, RL environments, and evaluations to encourage further open research and development within the AI community. This open approach is central to their strategy of making advanced AI accessible.

This release exemplifies Prime Intellect's strategic vision to empower a wider range of entities to engage in AI development, effectively enabling "every company to be an AI company." The company plans to continue scaling agentic RL and integrating a broader array of environments from its community-driven Environments Hub to further enhance future model capabilities.