Ollaya: pull and run open decision models like Ollama
A new project called Ollaya landed this week with a simple pitch: Ollama, but for decision models. One command (ollaya run laya) pulls a model and serves it on localhost:11435 with TypeSafe’s /v1/systemone shapes. The TypeSafe Python SDK works unchanged against the local server.
What it serves. Eight model families out of the box: Laya (Convai), Decider (Mapika), NLI (Moritz Laurer), GLiClass (Knowledgator), Qwen3Guard, Decision (vLLM Semantic Router), Kev (Jared Palmer) and Von (Victor Hugo Panisa). Sizes range from 322M to 7.9B. Laya and NLI run on the Apple GPU via MLX; the rest need an NVIDIA GPU (CUDA 13) or fall back to CPU.
Platforms. Desktop apps for macOS (Apple Silicon, 14+), Windows 10/11 and Linux x86-64 with systemd service support, plus a Docker image on GHCR. Linux ARM64 is CLI-only with no GPU acceleration.
Latency claims (UnverifiedClaim). On an RTX 4090 via the HTTP API, Ollaya reports medians of 8–10 ms for Laya, 15–20 ms for NLI/GLiClass, and 155–190 ms for Decider. For context, they cite third-party benchmarks showing the hosted Jev API at 236–276 ms — but the comparison mixes local GPU with network round-trip, so treat it as order-of-magnitude.
What it is not. Ollaya does not train any weights. It’s distribution and inference infrastructure, not a new decision model. The independence disclaimer is explicit: “not affiliated with Ollama or TypeSafe.”
The GitHub repo (Apache-2.0, Rust) had 446 stars and 16 forks three days after creation.
Runtime entry: /runtimes/ollaya. Related models: Laya, Decider, Kev.