What Magic does
Magic (Magic AI, Inc.) is a private San Francisco-based AI research and engineering company building long-context foundation models and agentic systems aimed at automating software engineering and AI research. The company says it pursues “frontier code models” by combining frontier-scale pre-training, domain-specific reinforcement learning, ultra-long context (trained and/or used during inference), and inference-time compute, with the stated goal of accelerating “safe AGI” via automated research and code generation. Based on its public technical updates, Magic’s differentiation centers on (1) ultra-long context mechanisms for code/repos, (2) compute-efficient pretraining and scaling research, and (3) model/service scaling using dedicated GPU clusters built in partnership with Google Cloud and NVIDIA.
Magic’s current “product” positioning is closely tied to its long-context model research; its blog describes named context windows and model efforts (LTM-1 and LTM-2) and a progression toward models that can reason over extremely large codebases. The company also publishes an AGI Readiness Policy and maintains a vulnerability disclosure program, indicating an intent to operate with explicit safety governance as capabilities improve.
From an organization standpoint, Magic appears to operate as a tightly staffed lab (public materials reference a small team) with emphasis on hiring for research, training/inference systems, supercomputing infrastructure, and security. Strategically, Magic’s public statements emphasize vertical integration (research, model training, and product/front-end iteration) and scaling compute via large GPU clusters, which supports both longer-context inference and larger-scale training runs.