THX-01

THX-01

Answers typed questions about a ticket, email or document in one fast pass, with calibrated probabilities, extracted numbers, excerpts and citations. A 322M multilingual model for classification and extraction.

Made by doofz on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: THX-01 is a multilingual decision model from HAL-X AI. You give it a piece of text, such as a message, ticket, email, document or JSON record, plus typed questions in plain language. All the questions are answered together in a single pass, which takes about 10 ms on one GPU. The answers come with calibrated probabilities, and it can also return numbers, exact excerpts and citations.

Who it suits: Developers who build ticket routing, document triage and data extraction, especially for Azerbaijani, Russian, English and Turkish text.

Industry | Who uses it | How they use it | Example
Customer support | Teams that route support tickets | Classify each ticket into a category with a probability | Sending a billing complaint to the billing queue
Finance | Analysts who read reports | Pull figures and quotes out of documents | Extracting a quarter's revenue with its supporting sentence
Automation | Developers of agent systems | Make fast yes or no and multiple choice decisions on agent traces | Checking whether an email needs a human reply

Skip it if you need free text generation, since it answers typed questions and does not write text.

Where to use it: It runs from Python through the thx01 library, and a REST server is available as an extra. The files download on Windows, macOS or Linux.

• Python, with the thx01 library from PyPI
• A GPU for the quoted speed, though the card says it uses a GPU if available
• A REST server, installed with the server extra

Use case | Who | What they get back
Ticket classification | Support teams | A category with a probability for each option
Number extraction | Analysts | A value taken from the document, or null if absent
Excerpt and citation | Developers | Exact text spans and the parts that support an answer

How to set it up: Install the library from PyPI. The weights download from this repository the first time you use it.

• Python with pip
• Enough memory for a 322M model, about 644 MB in F16

1. Install the thx01 library, with the server extra if you want the REST server.
2. Load the model by name in Python.
3. Send a document and a set of typed questions.

pip install thx01                 # library
pip install "thx01[server]"       # plus the REST server

You can also fetch the files yourself:

pip install -U huggingface_hub
hf download doofz/THX-01

How to use it: Load the model, pass a document and a dictionary of questions, each with a type, and read the answers. Mark a question with the cite option to get its supporting text.

import thx01

agent = thx01.load("doofz/THX-01")              # GPU if available

doc = ("Northwind Corp. reported Q3 2026 results. Revenue for the quarter was $48.3 million, up 14%. "
       "Net income was $6.2 million, compared with $4.1 million a year ago. "
       '"We will open two offices in Baku," said CEO Laura Chen.')

result = agent.decide(doc, {
    "kind":    {"type": "choice", "question": "What is this document?",
                "criteria": {"earnings": "earnings report", "complaint": "customer complaint", "invoice": "invoice"}},
    "revenue": {"type": "number", "question": "What was the revenue, in US dollars?", "cite": True},
    "quote":   {"type": "excerpt", "question": "What did the CEO say?"},
})
# kind -> earnings, revenue -> 48300000 (+ the supporting sentence),
# quote -> verbatim span starting "We will open two offices in Baku"

Question type | What it returns
choice | One of the options, with a probability for each
noul | The probability of yes
score | An ordinal level and its distribution
number | A numeric value taken from the text, or null when the text gives none
excerpt | An exact piece of the text, plus where it starts and ends

Example: You pass a short earnings report and ask three things: what kind of document it is, what the revenue was in US dollars, and what the CEO said. It answers earnings, 48300000 with the sentence that supports it, and the CEO's exact words.

For newcomers and experts: If you are new to this, think of it as a quick form filler. You give it a text and a few questions with fixed answer types, and it returns each answer with how sure it is.

If you know the field, it is an mmBERT-base encoder with a decision head, trained with reinforcement learning on a proper scoring rule so that honest probabilities score best. Number answers are copied from the text and never converted between units or currencies, and excerpts are cut from the document, so they cannot hold invented text.

Text and typed questions go in
Encoder reads each question and state
Decision head answers in one pass
Probabilities, numbers or excerpts return

Good to know: Input is limited to 1,024 tokens per question, and longer text is truncated. The card reports its own test on 2,843 support tickets in four languages.

Title: Card test, support ticket classification
Measure | THX-01
Average accuracy | 98.4
Calibration error | 0.003
Latency | about 10 ms

• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: the text and questions you pass in
• Not for: unit or currency conversion, since number questions return null in that case

On Hugging Face, 13 November 2026

• Task: text classification
• Library: thx01
• Parameters: 322M
• Weights: about 644 MB in F16; about 161 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 789
• Likes: 93
• License: apache-2.0
• Languages: az, en, ru, tr, de, fr, es, ar
• Based on: jhu-clsp/mmBERT-base
• Fine-tunes: 1
• Tags: thx01, safetensors, decision-model, calibration, reinforcement-learning, proper-scoring-rules, routing, extraction, citation, multilingual, azerbaijani, thx-01

Links

• Model: https://huggingface.co/doofz/THX-01
• Files: https://huggingface.co/doofz/THX-01/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download doofz/THX-01

Then follow the usage in its model card. Nothing in this kit runs by itself.