Two days ago, we released THX-01, an open-source decision model that outperforms Jev on several benchmarks and can even run locally on a smartphone. Today, anyone can use it through our free API.
Hi everyone! I'm Farid, Co-Founder of HAL-X AI.
A couple of days ago, our team open-sourced THX-01, a 322M-parameter multilingual decision model designed to make AI decision-making faster, cheaper, and more accessible.
Yesterday, we opened our hosted API to everyone. No API key, no sign-up, and no payment required.
And the response has been incredible!
Since launch, we've already processed:
We're excited to see developers experimenting with what a small, specialized decision model can do.
Most applications don't actually need an LLM to generate paragraphs of text.
Sometimes, you just need to answer a simple question:
Using a large generative model for these tasks can introduce unnecessary latency and cost.
THX-01 takes a different approach.
It's a non-autoregressive model that processes a document or message alongside typed questions and returns structured answers with calibrated probabilities in a single forward pass.
Approximately 10 milliseconds on our benchmark hardware.
No token-by-token generation.
You can make your first request right now.
No authentication required.
curl https://api.hal-x.ai/v1/systemone \
-H "Content-Type: application/json" \
-d '{
"state": "My card was charged twice for one order",
"questions": {
"team": {
"type": "choice",
"question": "Which team handles this?",
"criteria": {
"billing": "billing",
"tech": "technical",
"sales": "sales"
}
}
}
}'
THX-01 evaluates the available choices and returns the predicted decision with probabilities.
The API supports:
| Type | What it does |
|---|---|
| Choice | Classification with probabilities |
| Yes/No | Binary decisions |
| Score | Ordered scoring |
| Number | Extract exact numbers from documents |
| Excerpt | Extract verbatim passages |
| Citations | Identify supporting text |
You can submit multiple questions in a single request.
Full API documentation: https://api.hal-x.ai/docs/thx-01
We benchmarked THX-01 against TypeSafe Jev 1.13 using 2,843 multilingual support tickets across four languages.
Here are the results:
| Metric | THX-01 | Jev 1.13 |
|---|---|---|
| Accuracy | 98.4% | 97.4% |
| Calibration error | 0.003 | 0.007 |
| Measured latency | ~10 ms | 331 ms |
| Open weights | Yes | No |
THX-01 achieved higher accuracy and better calibration on this benchmark, with approximately 33x lower measured latency.
The latency measurements use different serving setups, so they're not a controlled hardware-to-hardware comparison.
We don't claim THX-01 outperforms Jev everywhere, but these results demonstrate how competitive specialized, lightweight decision models can be.
THX-01 contains just 322 million parameters.
It was designed to support efficient inference without requiring a massive GPU infrastructure.
With a suitable runtime, it can even run locally on smartphones.
You can also install the Python package:
pip install thx01
The model supports 18 post-training languages, with particular attention to Azerbaijani and multilingual robustness.
We're releasing THX-01 under Apache 2.0, including its model weights, source code, and training scripts.
We believe developers should have the freedom to run, modify, deploy, and build on top of their AI models.
At HAL-X AI, our goal is to contribute to the global open-source AI ecosystem and make efficient AI accessible to more people.
That's also why we've made the hosted API free to use, subject to rate limits.
Free API: https://api.hal-x.ai/docs/thx-01
Model weights and code: https://huggingface.co/doofz/THX-01
Interactive demo: https://huggingface.co/spaces/doofz/THX-01-demo
Python package:
pip install thx01
We're still improving THX-01, and we'd love to hear feedback from developers building AI agents, automation pipelines, document processing systems, and classification services.
If you build something using THX-01, we'd love to see it!