Skip to main content

Laplace 3B (i14) — not integrated

Laplace 3B is i14’s 3B decoder with gated linear attention (fixed ~29.5 MB recurrent state per sequence, no growing KV cache). AWS and i14 published Trainium2 training/serving results in October 2026.

OpenTryOn status​

PathStatus
Hosted APINot published
Local weightsNot released (contact i14 / AWS for evaluation)
RegistryNot added — nothing to call until weights or an endpoint ship

When Hugging Face weights and a vLLM/SGLang recipe appear, integrate as Path B (laplace-3b-local) on understand, mirroring Hy4/Sage local clients.

Do not confuse with unrelated Hugging Face org Laplaces-Red-Devils — that is a different team.