Ollaya runs Jev-style decision models locally and speaks TypeSafe's /v1/systemone API, so existing Jev clients switch with one env var

GitHub·high signalXBlueskyLinkedInCopy link

Ollaya (ollaya-dev/ollaya, Rust, Apache-2.0, created 2026-09-23, v0.7.1 today, 296 stars) reached 483 points on HN. It is an Ollama-style daemon for decision models, which return calibrated probabilities for typed choice/score questions and never generate text. It serves laya (ModernBERT-large, 421M), decider, NLI, GLiClass, Qwen3Guard, Kev and Von. The site claims 8-10 ms for five questions with laya on an RTX 4090, against 236-276 ms for hosted TypeSafe Jev. Setting TYPESAFE_BASE_URL=http://localhost:11435 points the official SDK at it, `ollaya mcp` exposes the models to Claude Code, and weights load from each author's Hugging Face repo pinned by sha256 rather than being re-hosted.