THX-02 is an open multimodal decision model by HAL-X. You give it a state (a message, a ticket, a document, a JSON record, an agent trace or an image) and typed questions in plain language. It returns an answer to every question with a calibrated probability for every option. It does not generate text, so it can never answer outside your options.

• #2 on the public Decision Index 0.3, ahead of 27B decision models, with 4.5B parameters
• Multimodal: text and up to 4 images per request (charts, receipts, forms, screenshots, photos)
• Multilingual: 20+ languages, including Arabic, Azerbaijani, Kazakh, Uzbek, Russian, Chinese and Japanese
• No training for new tasks: your options are part of the request
• Apache-2.0

Usage

ollama pull hal-x/thx-02

curl http://localhost:11434/v1/systemone -d '{
  "model": "hal-x/thx-02",
  "state": "AED 450 was charged twice at the same shop. Please refund the duplicate now.",
  "questions": {
    "team":   {"type": "choice", "instructions": "Which team should handle this?",
               "criteria": {"cards": "Card payments, disputes and refunds", "loans": "Loans and financing", "fraud": "Fraud and stolen cards"}},
    "refund": {"type": "noul", "instructions": "Does the customer ask for money back?"},
    "anger":  {"type": "score", "instructions": "How frustrated is the customer?", "criteria": ["calm", "annoyed", "angry"]}
  }
}'

Question types: choice (up to 255 options), noul (yes/no, returns the probability of yes) and score (an ordered scale of 2 to 10 levels). Images go in "images" as base64 strings.

Tags

Notes

This build answers directly. The self-gated thinking of the original model (a short thought before hard questions) is available in the PyTorch release and the HAL-X API: 63.8 without thinking and 65.06 with it on the Decision Index 0.3 sample.

• Model and docs: huggingface.co/HAL-X-AI/THX-02
• Free API: api.hal-x.ai/docs/thx-02
• Demo: huggingface.co/spaces/HAL-X-AI/thx-02