Questions about Ardana

What is Ardana?

Ardana is an open-source, self-hosted runtime for decision models. It is one binary that pulls a model straight from Hugging Face, runs it on your machine and serves it over HTTP. You give a model some text, called the state, and a few typed questions. It reads everything once and answers each question with a probability for every option. It never writes text, so there is nothing to parse and nothing to retry. The install line is on the home page.

How does Ardana relate to Ollama?

Ardana runs decision models the way Ollama runs chat models, with the same verbs: ardana pull, run and serve. It can also read a model that is already in your Ollama store, as in ardana pull ollama:llama3.2, which gets the stock chat profile. Ollama 0.35 added its own /v1/systemone endpoint on September 29, 2026, with the nimble and tev1 models, which Ardana does not read today. The comparison page puts Ardana, Ollama and llama.cpp side by side.

Does Ardana run on my own machine?

Yes. The binary, the models and the server all run locally. ardana serve listens on 127.0.0.1:8000. A model is downloaded once, the first time you pull or run it, and read from disk after that, so every answer comes from your own machine. ardana list lists the models on your machine and ardana rm removes one.

Do I need a GPU?

No. One cold run of a 2B model at Q4_K_M on an AMD Ryzen 7 3800X with no GPU answered a support ticket in 1.09 seconds, with the model load on top. The README puts the same answer at about a tenth of a second on an M1 MacBook once the model is on disk. For safetensors models, which Ardana runs with Candle, it computes with Metal on Apple silicon and on the CPU everywhere else; those builds have no CUDA backend.

Which models and formats can Ardana run?

Ardana runs models straight from Hugging Face, in the format they were published in. ardana pull hf.co/<org>/<repo> takes a repository’s GGUF files, which llama.cpp runs. Since Ardana 0.3.2 it can also take the original safetensors checkpoint, which Candle loads without a conversion step when it supports the architecture the checkpoint declares. It also reads local GGUF files and models already in your Ollama store. An instruct model is read zero-shot in its chat template, and the model library lists ready families you can pull by a short name.

What is a System 1 decision model?

A System 1 decision model reads a state and typed questions and returns calibrated probabilities from one forward pass. It does not generate an answer token by token. A choice question picks one of 2 to 255 named options, a score question picks one of 2 to 10 ordered levels, and a noul question gives the probability of yes. Sorting a support ticket into billing, technical or sales is a classification, and it takes one choice question. Ardana can ask these questions of an ordinary instruct model too, reading it zero-shot.

Does Ardana have an API and SDKs?

Yes. ardana serve exposes POST /v1/systemone on 127.0.0.1:8000, along with GET /v1/models and GET /health. A request names the model, the state and the questions. The response holds every answer with its probabilities and the token usage, where output_tokens is 0. The API works with the TypeSafe clients: point the Python package typesafe-sdk at your server, and the playground copies a request as TypeScript for @typesafe-ai/sdk. If a model is not on the machine yet, the first request that names it pulls it.

What is the license, and where is the source code?

Ardana is open source under the Apache-2.0 license, and the code is on GitHub. It is written in Rust, and llama.cpp runs the GGUF models. You can build it from source with cargo xtask build, which builds the playground and then the ardana binary. The models you run keep their own licenses, set by whoever published them.

Which operating systems does Ardana support?

macOS, Linux and Windows. On macOS and Linux one shell line installs it, and on Windows one PowerShell line does. Either way you get a single binary in ~/.local/bin with the same commands and the same server. On Apple silicon, safetensors models run on Metal; on the other systems they run on the CPU. The README’s quoted speed comes from an M1 MacBook, and the blog’s cold run came from an AMD Ryzen 7 3800X without a GPU.

How do I install and update Ardana?

On macOS or Linux, run curl -fsSL https://ardana.ai/install.sh | sh. On Windows, run irm https://ardana.ai/install.ps1 | iex in PowerShell. Both scripts come from the latest release on GitHub and put one binary in ~/.local/bin. To move to a newer release later, run ardana update. Then ardana pull hf.co/<org>/<repo> downloads a model from Hugging Face, and ardana run asks it a question straight from the terminal.

Can I try Ardana in the browser?

Yes. The playground on this site runs a small model inside your browser tab with ONNX on WebGPU, and falls back to WASM where WebGPU is missing. Nothing gets installed. The same playground ships with the binary: ardana serve opens it at http://127.0.0.1:8000, where you can try a state and questions, read every probability and copy the request as curl, Python or TypeScript.