Vincent Weisser is co-founder and chief executive of Prime Intellect, which is building an open superintelligence stack for training, adapting, evaluating, and deploying advanced AI models. He argues that access to model weights means little if the infrastructure for improving those models remains controlled by a handful of companies.
From decentralized science to open models
Weisser studied at CODE University of Applied Sciences, co-founded dex.blue, and helped develop decentralized-science initiatives around VitaDAO, including its Fast Grants Fellowship and the Longevity Prize. He subsequently co-founded Prime Intellect with Johannes Hagemann, applying distributed coordination to the computing and infrastructure required for advanced AI.
The company’s OpenDiLoCo project addressed a central obstacle: training models across geographically dispersed hardware without requiring constant, high-bandwidth communication. In November 2024, Weisser co-authored the release of INTELLECT-1, a 10-billion-parameter model trained collaboratively across five countries and three continents, with its training data, checkpoints, and framework released publicly.
INTELLECT-2, released in May 2025, extended that approach to asynchronous reinforcement learning across distributed computing providers. Weisser later co-authored the technical report for INTELLECT-3, a 106-billion-parameter mixture-of-experts model with 12 billion active parameters, developed using an open reinforcement-learning stack.
- Training environments are essential infrastructure. Prime Intellect’s Environments Hub gives developers a way to build and share task environments, evaluation criteria, and reward signals. Weisser argues that agents improve when they train against the particular software, financial workflows, or computer-use tasks they must actually perform.
- Production feedback loops create specialized intelligence. Prime Intellect Lab connects evaluation, reinforcement-learning training, deployment, and inference. Teams can improve models using evidence from real product usage instead of relying entirely on general-purpose frontier systems.
- Outcomes matter more than token volume. Prime Intellect’s work with Ramp on spreadsheet search illustrates Weisser’s emphasis on specialized models that improve task performance, speed, and cost. His practical benchmark is whether the value produced exceeds the cost of running the model.
- Open models improve through shared infrastructure. Through NVIDIA’s Nemotron Coalition, Prime Intellect contributes reinforcement-learning environments, post-training infrastructure, and evaluation tools. NVIDIA’s Nemotron and Arcee AI’s Trinity are external model families that developers can customize using that broader ecosystem.
Prime Intellect announced a $130 million Series A in July 2026. Weisser’s focus is on making advanced model development accessible to companies building agents around their own tasks, products, and production data.
