Chai

Chai is a social AI platform rather than an assistant: people arrive to talk with characters other users built, and to publish their own, under the company's own framing of a platform for AI characters and stories. It runs its own models instead of reselling a frontier API — the research site lists architectures from 6B up to a 235B mixture-of-experts with 22B active parameters, trained with custom reward setups including BO8 and DPO — which is what lets long, permissive conversational threads exist at a cost a per-token reseller could not survive. The mobile apps are the product, iOS and Android, with a web client advertised as coming. The vendor states over a million daily active users as of March 2025 and roughly $70M a year in revenue for 2026, which are its own numbers, not audited ones.

Where it loses: it is not useful for work. There is no knowledge base, no citation, no code execution, and the characters are entertainment, so anything you would verify in a professional assistant you cannot verify here. Store policies, not a published safety framework, decide most of what gets moderated. Where it wins is scale of companionship on a self-operated model stack. Two checks before you take it seriously: read the data terms before typing anything personal, since chats live in a consumer app with no admin export or audit trail; and if you are evaluating the technology rather than buying the product, the interesting finding is that a mid-size lab can run its own model line at this scale — that is the part worth understanding, not the character list.