Jeff-Gemma4-E2B Calibrated zero-shot choice classification in one forward pass

Model card Project

What this demo shows

Jeff-Gemma4-E2B is a small decision model. You describe a situation and list your own options in plain words; it returns a calibrated probability for each option from a single forward pass. This page is a static preview with the model’s benchmark results, an official gameplay demo, and a request builder for the local inference API.

Static preview

This Space does not run the 4.6B-parameter model in the browser. The Gradio inference app is included in this repository and is ready to deploy on a GPU-backed Space once hardware access is available.

Benchmark accuracy chart comparing Jeff models and Jev
Public benchmark accuracy from the official model card.

Gameplay demo

The model also powers real-time decisions in classic games. This official recording shows Jeff-Gemma4-E2B playing Pac-Man.

Jeff-Gemma4-E2B playing Pac-Man. Additional Doom and Frogger recordings are available in the project repository.

Build an inference request

Use the builder below to create the exact JSON payload accepted by the local /v1/systemone endpoint. 

{
  "model": "jeff-latest",
  "state": {
    "voice_transcript": "open the engagement letter",
    "current_screen": "Deal overview"
  },
  "questions": {
    "intent": {
      "type": "choice",
      "instructions": "Which of these does the user want?",
      "criteria": {
        "1": "Engagement letter",
        "2": "Inbox",
        "3": "Deal settings"
      }
    }
  }
}

Run the model locally

  1. Clone the official Jeff repository: git clone https://github.com/firelex/jeff
  2. Install dependencies with uv sync.
  3. Download the checkpoint: uv run hf download mstrasser/Jeff-Gemma4-E2B --local-dir Jeff-Gemma4-E2B
  4. Start the server: JEFF_CHECKPOINT=Jeff-Gemma4-E2B PORT=8765 uv run jeff-serve
  5. POST the generated JSON to http://localhost:8765/v1/systemone.