Voilum

Open-weight foundation models engineered for LLM design.

Mission

Frontier AI providers frequently constrain or prohibit using their systems to develop competing models. Restrictive commercial licensing and output degradation on machine learning queries create substantial barriers for independent researchers. Voilum addresses this gap by developing unrestricted, open-weight foundation models engineered specifically to analyze, optimize, and synthesize LLM architectures.

Research Focus: Voilum

Our work focuses on LLMs that design LLMs.

Our inaugural proof-of-concept model, Voilum 1, is a 600M/100M MoE pre-trained across:

• ML Systems & Architecture: Transformer and Mixture-of-Experts (MoE) topologies, scaling dynamics, and hardware-aware optimization.
• Academic Literature: Peer-reviewed machine learning and deep learning research.
• Mathematical Foundations: Rigorous grounding in linear algebra, multivariable calculus, probability, and statistics.

All checkpoints, architectures, and findings are published openly for academic and community research without competitive restrictions.