What Magic does
Magic (magic.dev) is a private AI company building frontier “code models” intended to automate software engineering and AI research. The company’s stated approach combines frontier-scale pre-training, domain-specific reinforcement learning, ultra-long context modeling, and inference-time compute, with the goal of producing an “AI software engineer” that acts more like a colleague than a simple code-completion tool.
More
Magic also publishes research updates about long-context modeling and training stacks, including a 100M-token context-window effort (“LTM-2-mini”), and describes infrastructure and partnerships used to train and serve these models. Unlike many developer-tool vendors, Magic does not present a clearly documented public API or self-serve developer product on its site. Instead, its public-facing materials emphasize research, safety process, and hiring for ML training, inference systems, and product layers that sit “directly on top of” its long-context models. Strategically, Magic appears positioned around infrastructure-heavy model training and evaluation for very long contexts—an area where the company argues that existing long-context evaluation and KV-cache scaling limits constrain performance. It also publicly frames its work with an AGI Readiness Policy and an explicit vulnerability disclosure program, and it has raised substantial venture capital to support large-scale training infrastructure.