Ollaya packages open decision models behind a local server that speaks TypeSafe's API, and reached 519 points on Hacker News in a day

Anyone paying per call to route tickets, tag emails, or score content by intent can now do the same job on a laptop by changing one environment variable. Ollaya, released on 23 September 2026 under Apache-2.0, pulls open decision models from Hugging Face and serves them behind an HTTP API that matches TypeSafe's Jev. A Show HN thread reached 519 points in a day and the repository at ollaya-dev/ollaya carries 338 stars three days after creation.

What it does

A decision model answers a typed question over a fixed schema and returns a probability, not free text. Ollaya packages several open ones together and serves them through the same /v1/systemone and /v1/models endpoints that TypeSafe uses for Jev. The site states the official TypeSafe Python SDK 0.7.1 works unchanged against a local server after TYPESAFE_BASE_URL is set to http://localhost:11435. That is the whole switching cost for a team already writing against Jev.

The models included cover common classification shapes: Laya for choice, score and yes-or-no in one forward pass across more than 100 languages, Decider on Qwen3.5 for higher accuracy at a longer latency, NLI zero-shot classifiers by Moritz Laurer, GliClass for many-option questions where cost stays flat, Qwen3Guard for safety triage in 119 languages, and Von for inputs up to 8,192 tokens. Weights come from each author's Hugging Face repository, pinned to a commit and checked against a sha256 hash. Ollaya says it never re-hosts them.

What the numbers say

Ollaya publishes its own latency figures for a five-question request through the HTTP API on an RTX 4090: 8.1 milliseconds for the multilingual Laya, 9.6 for the English Laya, 14.7 for GliClass, 20.4 for NLI, 155 for the 0.8-billion-parameter Decider and 190 for the 2-billion. It compares those to a range of 236 to 276 milliseconds for TypeSafe's hosted Jev API. The Jev figures come from two third-party benchmarks Ollaya names as AbdelStark/jev-benchmarks and nibzard/decision-model-benchmark, and the local numbers include no network. The page states the setups differ and to read the comparison "as an order-of-magnitude comparison" rather than a like-for-like test.

The example on the site sends a 34-word tool call to decider:2b and gets back four typed answers, one of them a 0-to-2 risk score with a written reason, in 178 milliseconds on the same GPU.

Where it runs

Ollaya has a desktop app and a command line for macOS, Windows and Linux, plus a Docker image on GHCR. Every model runs on the CPU. An NVIDIA GPU on Linux, Windows, WSL 2 or Docker with driver R580 or newer and CUDA 13 takes a request down to milliseconds. On Apple Silicon, Laya and NLI use the Apple GPU through MLX; other models fall back to the CPU. The server binds to 127.0.0.1 by default, so a request never leaves the machine unless a team makes it.

The value here is the shape of the substitution. A team pointing a Jev client at a local Ollaya server keeps the same request and response contract, so the code around it does not change, and the data stays on the machine that already holds it. The vendor comparison should be run on real workloads before any budget move, since every Ollaya number came from Ollaya.

Source

• Ollaya: ollaya.dev
• GitHub: ollaya-dev/ollaya
• Discussion: Hacker News thread