TraceWeave
Category: Foundation Models / LLMs
Korean on-device AI startup building Phase-TITAN, a post-Transformer architecture with fixed-size state memory (no KV cache) that lets AI agents run locally on smartphones, vehicles, robots, drones and smart glasses. TraceWeave was founded in 2025. The company is led by Jaeyoung Park (박재영). Based in Seoul, South Korea. Team size: 1-10. Total funding raised: $2.3M. Latest round: Seed. Key investors include Mirae Asset Capital, Quad Ventures, NAVER D2SF, Antler Korea.
- Founded
- 2025
- Headquarters
- Seoul, South Korea
- Team size
- 1-10
- Total funding
- $2.3M
Value proposition
Removes the Transformer KV-cache memory bottleneck so that long-context, multi-step tool-using AI agents can run fully on memory-, battery- and heat-constrained devices without cloud calls. The company says its 1B Phase-TITAN model holds about 0.04 GB of state at 1M-token context, versus tens of GB for a comparable Transformer's KV cache.
Products and solutions
Phase-TITAN (PT) post-Transformer architecture, PT Language Model (v5 1B with ~40 MB per-sequence state, v3.5 3B demoed on a Galaxy S26 Ultra), aimed at on-device, long-context, tool-use agents, PT Training-Free Dynamics Monitor for online anomaly/fault detection on vehicles and UAVs, with a heap-free Rust edge core, PT World Model, a persistent interactive video world model for robotics/AV simulation, and in-vehicle AI agents that run on a single CPU core with under 1 GB of memory in internal tests.
Unique value
A proprietary attention-free, KV-cache-free architecture built from scratch rather than on open-source models. Memory stays constant regardless of context length while the company reports Transformer-class performance on RULER (0.819 at 128K) and BFCL-v3 (0.5147).
Target customer
Automotive OEMs and their autonomous-driving affiliates, automotive electronics/software vendors, and robot, drone, smart-glasses, wearable and smartphone makers needing on-device AI (B2B, API/SDK).
Industries served
Automotive / autonomous driving, robotics, drones / aerospace, smart glasses / AR / wearables, mobile devices, industrial machine monitoring
Technology advantage
Architecture is grounded in computational neuroscience and physics: intelligence is represented as evolving compact state rather than a growing context. The team owns the full stack, from architecture design and pretraining to the inference engine. One model family covers language agents, training-free dynamics monitoring and world models. It has commercial validation through a paid automotive PoC.
How they differentiate
TraceWeave replaces the Transformer outright with a fixed-state architecture instead of compressing or quantizing existing models (as Nota does). Its focus is agentic tool use and physical-edge use cases (vehicles, UAVs, world models) rather than general-purpose edge LLMs or voice. It also offers a training-free dynamics monitor that needs no labeled failure data.
Main competitors
Liquid AI (US, non-Transformer Liquid Foundation Models for edge), Cartesia (US, state-space-model architecture), Nota AI (KR, on-device model compression/optimization)
Key partnerships
Paid PoC completed with the autonomous-driving affiliate of a global automotive group (driving-data frame selection without human annotation), in-vehicle agent validation ongoing with a US/global automotive electronics software company, NAVER D2SF portfolio (resident at D2SF @Gangnam), Antler Korea company-building program (cohort 7), NVIDIA Inception Grand Challenge 2026 Top 20 finalist
Notable customers
Undisclosed autonomous-driving affiliate of a global automotive group (paid PoC), undisclosed global/US automotive electronics software company (validation)
Major milestones
2025-10: Incorporated via Antler Korea Startup Generator. 2026-03: Antler Korea cohort 7 investment. 2026-07: Filmed a 33-minute persistent PT World Model demo. 2026-08: Named a Top 20 finalist in the NVIDIA Inception Grand Challenge. 2026-09: Selected for MSS 'K-Startup' AI league. 2026-10: Raised a ₩3B seed led by Mirae Asset Capital, with NAVER D2SF as a new investor.
Growth metrics
3 employees as of Aug/Sep 2026 (national pension data). The seed valued the company at about ₩30B (~$22M), reported as pre-money in the PR Newswire release, about one year after founding.
Market positioning
Early-stage Korean deep-tech challenger in post-Transformer / on-device agent architectures, targeting automotive and physical-edge AI first, then digital edge (smartphones, AR glasses, wearables).
Geographic focus
South Korea base with global automotive and device customers (US automotive software partner, global automotive group)
About Jaeyoung Park (박재영)
Ex-Google corporate strategy; Ex-42dot (Hyundai Motor Group autonomous driving/software affiliate) corporate strategy; 12+ years in AI and technology. Co-founded TraceWeave via the Antler Korea Startup Generator in Oct 2025.
Latest news about TraceWeave
More Foundation Models / LLMs companies
Official website: https://traceweave.net