Research Engineer, Long Context

Magic is building AI with very long context windows and we're looking for a Research Engineer to push the boundaries of context length in language models. You will work on novel attention mechanisms, memory architectures, and training techniques that enable models to process millions of tokens. This is a frontier research role working on one of the most challenging problems in AI. You will develop new approaches to efficient attention, position encoding, and memory management for extremely long sequences. The ideal candidate has deep expertise in transformer architectures and systems optimization.

- 5+ years of ML research/engineering experience - Deep understanding of attention mechanisms and transformer architectures - Experience with long-context methods (sparse attention, linear attention, etc.) - Expert PyTorch and CUDA skills - Experience with distributed training at scale - PhD in ML/CS preferred