One foundation, two products.
Security and memory live inside the representation layer — not in a filter bolted on afterwards.
5,059,040 parameters · Trained on CPU · Research License v1.0
What DaoTi is
The technical layer beneath the products.The technical layer beneath the products.
It starts at the representation layer
It starts at the representation layer
In Kanhe: the decision happens before the request leaves the machine, not at a gateway afterwards.
Directions can be separated
Directions can be separated
In Loong Recall: memory runs on its own direction and never shares a path with security decisions.
Platform-level reuse
Platform-level reuse
Both products share one premise: once the base is frozen, capability grows through adapters.
Products grown from the foundation
Two lines, each evolving independently.Two lines, each evolving independently.
Verifiable technical facts
Numbers come from public experiment logs and can be reproduced.Numbers come from public experiment logs and can be reproduced.
Retrieval loss is almost completely flat along that direction.
Representations from different domains are highly isomorphic along text_proj (CKA).
That many parameters, all trained on consumer-grade CPUs.
Start from the foundation
Weights and inference code are public. Kanhe and Loong Recall are shipping one after another.