The conversational role-play agent is based on Sophon, a character from The Three-Body Problem universe, and is positioned around multi-turn interaction and memory that persists across sessions.
Infinigence provides the development and serving stack through its Agentic Infra offering and AgentWorks platform. The companies say the system turns source material into callable, traceable and reusable runtime resources for character knowledge, expression and interaction memory. Its response path separates character expression, knowledge retrieval and content checking, then uses a generate, judge and targeted-revision loop.
For serving, Infinigence says its PDD cross-cluster heterogeneous inference design splits work into Prefill, Relay Decode and Main Decode stages. The company reports a 51.5% reduction in time to first token and a 37.5% reduction in per-token cost for the deployment.
The source does not state which model powers Sophon Agent 1.0, whether the agent is publicly available, its price, workload scale, benchmark methodology, or an independent measurement of the performance claims.