No AI release moved the market like DeepSeek’s in January 2025: near-frontier results at a fraction of the assumed training cost, with the weights published and the licences permissive. It settled one argument for good. Open-weight models are a serious business option, and the rest of the industry had to justify its prices.

Three ways to use DeepSeek

The name covers three quite different things, and only one suits client data:

  • The open-weight models. The ones that matter most for firms: downloadable, MIT-style licensed, and served by many third-party hosts. Used this way, your privacy depends on the host you choose, not on DeepSeek.
  • The direct API. Priced aggressively, with data processed by DeepSeek under its own terms. Check it with the standard checklist, like any processor.
  • The consumer app. It briefly topped the download charts, and it goes on the same approved-or-not list as every other consumer AI app in your tools policy.

The business case for DeepSeek

The case is capability for the money, and it’s real. For extraction, classification, drafting and reasoning-heavy back-office work, DeepSeek-class models through a host you choose give you most of what the frontier models do at a fraction of the cost. That’s the open-weight argument in one example.

The points against are just as real, and there’s nothing awkward about naming them:

  • Provenance matters to some boards and some clients, and pretending otherwise helps nobody. Put the question to your risk owner and record the answer.
  • Polish is thinner than the frontier products: tooling, enterprise contracts and depth of support.
  • The route in for a regulated firm is through hosts and terms you can put on paper, not the consumer app. The method from this site’s benchmark applies unchanged: five of your own tasks, through a host, judged on the output.

The decision is the same as any open-weight decision, with the provenance question added.