Today, we're introducing THX-01, an open-source, non-autoregressive decision model developed by HAL-X AI in Baku, Azerbaijan.
With just 322 million parameters, THX-01 is lightweight enough for deployment on CPUs and potentially smartphones, without requiring expensive GPU infrastructure.
Unlike traditional large language models, THX-01 doesn't generate text token by token. Instead, it processes inputs and returns structured decisions with calibrated confidence scores in a single forward pass.
Key Highlights
- 322M parameters, built on mmBERT with a specialized decision head.
- ~10 ms inference per request on a GPU.
- CPU-compatible, with a compact architecture suitable for mobile and edge deployment.
- 98.4% accuracy on our multilingual support-ticket classification benchmark, compared to 98.5% for Claude Sonnet 5.5.
- 18 languages, including Azerbaijani, English, Russian, and Turkish.
- Calibrated confidence scores, allowing applications to make decisions based on estimated uncertainty.
- Apache 2.0 license, with publicly available model weights.
Why We Built It
Not every AI task requires a massive generative model.
Classifying messages, routing customer requests, detecting spam, extracting document values, and making structured decisions can often be handled by significantly smaller models.
THX-01 was designed specifically for these workloads, prioritizing speed, efficiency, and reliability.
Our goal is to make intelligent decision-making accessible across environments, from production servers to resource-constrained edge devices.
Get Started
Install THX-01 using pip:
pip install thx01
Load the model:
import thx01
agent = thx01.load("doofz/THX-01")
Open Source
THX-01 is available under the Apache 2.0 license.
Model and documentation: https://huggingface.co/doofz/THX-01
Built by HAL-X AI, Baku, Azerbaijan.
We believe the future of AI isn't only about building larger models. It's also about building smaller, specialized models that make powerful capabilities accessible to everyone.
Top comments (0)