THX-02: A decision model that assigns probabilities to multiple-choice questions in tens of milliseconds

HAL-X-AI's "Thinking Decision Model": Given a state (messages, tickets, documents, JSON records, agent trajectories, photos, or charts) and a question written in natural language, it outputs probabilities for each option of every question without generating free text. Questions are categorized into three types: single-choice (up to 255 options), true/false, and ordinal scoring; a single request can include up to 4 images. The model has 4.5B parameters, along with a 60M thinking adapter: if the highest probability of the direct answer falls below a threshold and internal judgment requires reasoning, it first writes a few dozen words of reasoning notes before deciding. The authors report median latency per request of 50 milliseconds without thinking and 79 milliseconds with automatic thinking (on one RTX PRO 6000), achieving 65.06 on a sample from the Decision Index 0.3 public set.

Original title: HAL-X-AI/THX-02

HAL-X-AI's "Thinking Decision Model": Given a state (messages, tickets, documents, JSON records, agent trajectories, photos, or charts) and a question written in natural language, it outputs probabilities for each option of every question without generating free text. Questions are categorized into three types: single-choice (up to 255 options), true/false, and ordinal scoring; a single request can include up to 4 images. The model has 4.5B parameters, along with a 60M thinking adapter: if the highest probability of the direct answer falls below a threshold and internal judgment requires reasoning, it first writes a few dozen words of reasoning notes before deciding. The authors report median latency per request of 50 milliseconds without thinking and 79 milliseconds with automatic thinking (on one RTX PRO 6000), achieving 65.06 on a sample from the Decision Index 0.3 public set.

Why it matters For judgments such as ticket routing, content moderation, and UI operations, there is no need to retrain for new tasks; labels are specified directly in the request.

Upvotes 48 · Likes 50 · Downloads 2134

Reasons

  • Demonstrates uses such as multilingual routing of customer service tickets, booking flights by clicking buttons step-by-step based on screenshots, and playing Snake based on visual input, none of which required task-specific training.
  • Provides a dependency-free PyPI client and self-hosted server, supporting batch interfaces; the API can be called on Android Termux, and it can run locally on phones with sufficient memory.
  • Weights are licensed under Apache-2.0.

Usage

  • After running pip install "thx02[local]", use from thx02 import THX02; THX02.load("HAL-X-AI/THX-02"), then call model.decide(state, question_dict) to obtain probabilities for each option.
  • Self-hosted service: run thx02-serve --model . --port 8200, and send requests to /v1/decide or /v1/decide/batch; set think to "auto" to enable automatic thinking.

Risks

  • Scores are self-tested by the authors on a sample of 6,018 requests; full public set and private leaderboard tests have not yet been run. Scores for other models in comparison rows are taken from their respective published results.
  • The thinking feature requires a separate GPU inference service; CPU and mobile devices can only provide direct answers.
  • The authors caution that probabilities are calibrated only on the training distribution; when switching to significantly different domains, verification with several hundred labeled samples is required first.
  • This entry has not been run; refer to the model card for actual performance.

Open

Hugging Face Trending

Explore the latest AI Radar updates