HAL-X AI has officially released THX-01, a 322-million-parameter multilingual decision model, and made its hosted API completely free to use. The release, announced by Co-Founder Farid on October 10, 2026, positions the model as a lightweight alternative to large generative models for structured decision-making tasks. Since the open-source launch two days prior, the team reports processing over 25 million tokens and handling 211,000 decisions, signaling strong early developer interest in specialized, low-latency AI components.

Non-Autoregressive Architecture for Speed

Unlike standard LLMs that generate text token-by-token, THX-01 utilizes a non-autoregressive architecture designed for single-pass inference. The model processes a document alongside typed questions and returns structured answers with calibrated probabilities in approximately 10 milliseconds on benchmark hardware. This approach targets specific utility tasksβ€”such as spam detection, support ticket routing, and invoice extractionβ€”where the latency and cost overhead of generative models are often unnecessary for simple classification or extraction needs.

Benchmarking Against Jev 1.13

HAL-X AI compared THX-01 against TypeSafe Jev 1.13 using a dataset of 2,843 multilingual support tickets across four languages. The results show THX-01 achieving 98.4% accuracy compared to Jev's 97.4%, with a calibration error of 0.003 versus 0.007. Most notably, THX-01 demonstrated a measured latency of ~10 ms, roughly 33 times faster than Jev's 331 ms. The company explicitly notes that these latency figures come from different serving setups and are not a controlled hardware-to-hardware comparison, framing the results as proof of competitive viability rather than absolute superiority.

Open Weights and Mobile Capability

The model is released under the Apache 2.0 license, including all model weights, source code, and training scripts. With only 322 million parameters, THX-01 is designed to run efficiently without massive GPU infrastructure, with claims that it can operate locally on smartphones via suitable runtimes. Developers can install the Python package via pip install thx01 or access the model weights on Hugging Face. The API supports various decision types, including choice classification, binary yes/no decisions, ordered scoring, and verbatim excerpt extraction, all accessible without an API key or sign-up.

Key Takeaways

  • THX-01 is a 322M-parameter open-source model optimized for structured decision tasks, not text generation.
  • The free API requires no authentication, lowering the barrier for developers testing specialized AI components.
  • Benchmarks show a 33x latency advantage over Jev 1.13, though hardware comparisons remain uncontrolled.
  • The model supports 18 post-training languages and is licensed under Apache 2.0 for commercial and personal use.

The Bottom Line

THX-01 proves that not every AI problem requires a generative model; for structured decisions, specialized, lightweight architectures offer superior speed and accessibility.