Ollajev

Like Ollama, for decision models.
Run System One decision models from Hugging Face on your machine, behind the same /v1/systemone API as TypeSafe's Jev.

Product page · Install · Quick start · Models · API

• No generated text. Send one state and any number of typed questions; get a calibrated probability for each, in one forward pass.
• Images, audio and video beside the state, with Clef-Omni, d1-omni, and Clef on Apple Silicon through MLX.
• Ollama-style workflow: pull, list, run, serve, ps, rm, plus a terminal model manager and a browser playground.
• Drop-in for typesafe-sdk: point TYPESAFE_BASE_URL at it; same routes, same request and response shapes.

Install

curl -fsSL https://raw.githubusercontent.com/nvkudva/ollajev/main/install.sh | sh   # macOS, Linux
brew install nvkudva/tap/ollajev                                                    # Homebrew
uv tool install ollajev                                                             # PyPI

Windows (PowerShell): irm https://raw.githubusercontent.com/nvkudva/ollajev/main/install.ps1 | iex

• llama.cpp is compiled during the install: macOS needs the Xcode Command Line Tools (xcode-select --install), Linux a C/C++ compiler (sudo apt install build-essential). The script checks both.
• Add -s -- --service to the script (or run brew services start ollajev) to keep the server running at login.
• With uv on Linux, add --index https://download.pytorch.org/whl/cpu, or PyPI's PyTorch brings several GB of CUDA libraries.

curl -fsSL https://raw.githubusercontent.com/nvkudva/ollajev/main/install.sh | sh -s -- --uninstall

Quick start

ollajev           # first run opens the model manager: pick a model, Enter to download, s to serve
ollajev serve     # serve the default model and open the playground

The playground at http://127.0.0.1:8000/playground lets you pick or download a model, write a state and questions, attach an image, audio clip or video, and compare answers across models.

Example

curl -s http://127.0.0.1:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": "I was charged twice for the same order and nobody answers my emails. I want my money back now.",
  "questions": {
    "area":    {"type": "choice", "instructions": "Which product area is this about?",
                "criteria": {"refund & dispute": "A billing dispute or refund request", "card": "Anything about a card",
                             "other": null}},
    "urgency": {"type": "score", "instructions": "How urgent is this message?",
                "criteria": ["Can wait", "Needs attention this week", "Needs attention today"]},
    "refund":  {"type": "noul", "instructions": "The customer is asking for a refund."}
  }
}'

{
  "model": "Mapika/decider-4b-GGUF:Q4_K_M",
  "answers": {
    "area":    {"type": "choice", "choice": "refund & dispute", "confidence": 0.9642,
                "probabilities": {"refund & dispute": 0.9761, "card": 0.0104, "other": 0.0135}},
    "urgency": {"type": "score", "score": 1.712, "confidence": 0.568,
                "probabilities": {"0": 0.0099, "1": 0.2682, "2": 0.7219}},
    "refund":  {"type": "noul", "noul": 0.9433, "confidence": 0.8866}
  },
  "usage": {"input_tokens": 198, "output_tokens": 0}
}

About 0.4 s on an M3 Max. Question types: noul (yes/no), choice (pick one), score (expected level on an ordered scale). Leave model out to use the default; send "model": "<name>" to use another.

Images, audio and video

Add images, audio or videos as base64 data URLs:

curl -s http://127.0.0.1:8000/v1/systemone -H 'content-type: application/json' -d '{
  "model": "LiquidAI/d1-omni-600M",
  "state": "",
  "images": ["data:image/jpeg;base64,'"$(base64 < cats.jpg | tr -d '\n')"'"],
  "questions": {"cats": {"type": "choice", "instructions": "How many cats are there?",
                         "criteria": {"one": "One", "two": "Two", "more": "Three or more"}}}
}'

GET /v1/models lists what each model reads in limits.inputs. URLs and file paths are refused, so a request never makes the server fetch or read anything. Video is sampled at 2 fps and capped at 5 minutes.

Models

A model name is its Hugging Face repo id. GGUF repos take a quant tag (user/repo:Q8_0; Q4_K_M by default).

Any other repo from these families works too, such as a fine-tune: press a in the manager or ollajev pull user/repo. ollajev show <model> prints a model's limits; requests over them get a 422.

Commands

Run ollajev <command> --help for options. Models load on first request and unload after OLLAJEV_KEEP_ALIVE.

API

Errors are {"detail": [{"loc", "msg", "type"}]}: 404 model not downloaded, 403 not trusted, 422 invalid or over the model's limits, 503 not enough memory. Confidence uses TypeSafe's formulas, so it means the same thing for every model.

Configuration

Settings (o in the manager) are saved to ~/.ollajev/config.json; environment variables override them.

Security

• Local by default. The server binds to 127.0.0.1 and checks the Host header against DNS rebinding. Listening elsewhere requires OLLAJEV_API_KEY; put TLS in front before exposing it beyond a LAN.
• Repo code needs your trust. Julia, open-jev, Intern-Decision, Decision-1.0 and d1 run Python from their repo. The first pull shows the commit and asks you to trust it; trust is CLI-only, never over HTTP. Every repo is pinned to its first-downloaded commit.
• See SECURITY.md.

Known limits

• Requests to one model run one at a time (concurrent forwards crash on Apple GPUs).
• Before a load, free memory is checked against the weights plus 10%; a very long request can still run out.
• Cloudflare/clef-omni needs about 64 GB of GPU memory; on a Mac use mlx-community/clef-omni-4bit. The 8-bit MLX copies were not tested.
• d1-omni-600M is under the LFM Open License v1.0: commercial use only for organisations under 10M USD annual revenue.

Development

uv sync && uv run pytest

See CONTRIBUTING.md and the Changelog.

License

Apache-2.0. Vendored loaders keep their own licenses; see NOTICE. Ollajev is an independent project, not affiliated with Ollama or TypeSafe.