TraceWeave raises 3 billion won to develop AI agents that run on devices
The AMW Read
TraceWeave adds an early model-architecture entrant targeting on-device agents, but its seed financing, demonstrations and paid proof of concept do not establish broader technical or commercial impact.
TraceWeave raises 3 billion won to develop AI agents that run on devices
TraceWeave has raised approximately 3 billion won in seed funding led by Mirae Asset Capital, with Naver D2SF joining as a new investor, at a reported valuation of approximately 30 billion won. Founded in October 2025 by researchers from ETH Zurich, Google and KAIST, the startup is developing its own models using Phase-TITAN, an architecture designed to replace transformers. The company plans to use the funding to recruit talent and expand its business.
The architectural claim matters because longer context increases transformer KV-cache memory requirements, creating constraints for devices with limited memory, battery capacity and thermal headroom. TraceWeave says Phase-TITAN keeps memory usage constant as input length grows, supporting agents that call tools, interpret results and choose subsequent actions locally. This positions the company in the model-development layer of the agent market, where architecture could influence which workloads can operate without cloud connectivity. The reported smartphone CPU demonstration is an early signal; the article provides no comparative benchmarks for accuracy, latency or energy consumption.
For builders and investors, the concrete diligence question is whether those memory benefits hold alongside reliable task completion on customer hardware. TraceWeave has completed a paid driving-data analysis proof of concept with an autonomous-driving affiliate of a global automaker group and is validating vehicle-agent performance with an automotive software company. These engagements provide a route to commercial evaluation, but neither partner is named and the article reports no production deployment. Automotive validation is therefore the clearest disclosed test of whether the architecture can move beyond demonstrations.