Magic

Coding-model research with a public evaluation tool

Magic is a model-research company focused on software development, long-context systems and training efficiency. Its front door identifies the business as Magic AI, Inc. The research objective is broader than a released coding assistant, so product availability should be established from an access announcement rather than inferred from its ambitions or funding. [1][2]

Magic released HashHop on August 29, 2024. The benchmark uses randomly generated hash associations and multi-step retrieval to test long-context architectures while reducing semantic hints. The public implementation is available under MIT, and its version 1.0.0 release has an actual publication timestamp on that day. [2][3][4]

The same update described internally trained LTM-2-mini and prototype coding experiments. Publishing an evaluation tool did not release the model weights or make the internal system generally available. HashHop is a concrete public contribution whose existence can be verified without adopting the company's claims about model scale or efficiency. [2]

Magic's September 2026 update continued the training-efficiency research, describing held-out loss evaluations and compute scaling. Those comparisons and projected costs are the company's own measurements. They supply evidence of the research direction, without independently proving a production coding product or a public model release. [5]