Magic

Magic (magic.dev) is a privately held AI company building long-context models and agentic systems aimed at automating software engineering and, more broadly, “AI research and code generation” as a path toward safe AGI. The company’s approach emphasizes (1) ultra-long context windows for code understanding, (2) domain-specific reinforcement learning, and (3) inference-time compute, with the explicit goal of improving reliability and usefulness of generated code and engineering assistance. Magic positions its work as “science via product,” describing a vertically integrated stack that spans foundation-model training and user-facing systems. On its website and in research blogs, Magic highlights large-scale compute infrastructure (for example, thousands of NVIDIA GB200/H100-class GPUs) to train and deploy long-context models such as its LTM-1 line (introduced with a multi-million token context window) and research toward much larger (100M token) context modeling focused on software development workflows.