TraceWeave raises ₩3 billion to build AI agents that run without cloud calls

TraceWeave has raised approximately ₩3 billion in seed funding led by Mirae Asset Capital, with NAVER D2SF participating. The startup is developing AI agents that work entirely on devices rather than calling cloud servers. Its core claim: longer tasks need not bring the growing memory demands that make such agents difficult to run locally.

NAVER D2SF’s October 20 announcement put TraceWeave’s pre-money valuation—the value before the new investment—at approximately ₩30 billion. Seoul Economic Daily reported the funding as roughly $2.1 million. TraceWeave plans to use the proceeds to add research and engineering staff, develop its next-generation architecture and grow the business.

The company was co-founded in October 2025 by researchers with backgrounds at ETH Zurich, Google and KAIST. Led by co-founder and CEO Jaeyoung Park, it is based at NAVER D2SF’s Gangnam startup space. It develops its own models rather than building on existing open-source models.

Replacing the architecture, not just shrinking the model

TraceWeave calls its architecture Phase-TITAN. It is designed to replace the conventional Transformer structure, rather than simply fit an existing model onto smaller hardware. The target is a memory bottleneck: in Transformer-based models, the key-value cache, or KV cache, grows as the context being processed gets longer.

That growing cache adds to the demands on devices already constrained by memory, battery life and heat. TraceWeave says Phase-TITAN keeps memory use constant as context grows while supporting long-context understanding and tasks that require several steps. The intended agent can call tools, interpret their results and choose what to do next without contacting a cloud server.

In its memory comparison at a one-million-token context, without quantization, TraceWeave reports 0.04 GB for its one-billion-parameter Phase-TITAN model. It contrasts that with tens of gigabytes or more for a comparable Transformer’s KV cache. Tokens are the units used to measure context here; the comparison concerns the memory burden of processing that context.

Capability is the other half of the claim. TraceWeave says internal tests on the RULER and BFCL benchmarks put Phase-TITAN on par with Transformer models of the same size and larger. Those are company-reported results from internal testing, not an independent performance assessment.

Taking the KV cache out was never the hard part. Staying capable without it was.

Jaeyoung Park, TraceWeave co-founder and CEO

A phone demonstration and an automotive customer

According to NAVER D2SF’s announcement, TraceWeave has demonstrated a multi-step agent running solely on a smartphone’s central processing unit, or CPU, without cloud calls. Its commercial work also includes a completed paid pilot with the autonomous-driving affiliate of an unnamed global automotive group.

That pilot focused on driving-data analysis. The company says it identified frames that improve the partner’s autonomous-driving model without human annotation. A separate effort is validating in-vehicle agent performance with a global automotive electronics software company. The announcement does not name either partner.

Beyond automotive work, TraceWeave is exploring applications in robots, drones, smart glasses and wearables. Those remain commercialization opportunities rather than announced deployments. The funding therefore supports both further architecture research and efforts to turn its device demonstrations into a broader business.

NAVER D2SF head Sanghwan Yang framed the investment around bringing complex AI to devices beyond PCs. He also cited lower cloud-call costs and privacy protection as advantages of on-device processing. That is the investor’s rationale for backing TraceWeave, alongside its assessment that the startup can overcome existing models’ hardware constraints.