THX-01: An Open-Source Decision Model That Runs on Your Phone and Beats Jev

HAL-X AI released THX-01, a 322M-parameter open-source non-autoregressive decision model that answers typed questions over documents in a single forward pass, with roughly 10 ms GPU inference and support for 20 languages. On the company's internal four-language support-ticket classification benchmark, THX-01 reached 98.4% accuracy versus 98.5% for Claude Sonnet 5.5, with about 150× faster inference in its tests, and its calibrated confidence let a selective automation setup handle roughly 75% of tickets without errors. The model weights and inference tooling are available under the Apache 2.0 license via a pip install and Hugging Face.

Today, we’re releasing THX-01 , a new open-source decision model developed by HAL-X AI. Most production AI tasks don’t actually require generating text. They require making decisions: classifying support tickets, detecting phishing emails, routing requests, extracting invoice totals, or finding evidence in documents. Yet we often use large language models for these relatively simple operations. They’re powerful, but expensive, slower than necessary, and their reported confidence isn’t always meaningful. We built THX-01 to change that. A Different Approach to AI THX-01 is a 322M-parameter, non-autoregressive decision model designed to answer structured questions directly, without generating tokens one by one. It processes a document or message alongside typed questions and produces answers in a single forward pass. The result is approximately 10 ms inference on GPU , with a compact architecture small enough to run locally, even on a smartphone. THX-01 supports 20 languages , making it suitable for multilingual workflows and applications. Going Beyond Jev THX-01 follows a similar decision-oriented philosophy to Jev, but extends its capabilities. In our benchmarks, THX-01 outperforms Jev on several tasks while introducing capabilities that Jev doesn’t support natively. These include: - Native number extraction: Extracts exact numerical values from documents, including normalized formats. - Verbatim excerpts: Returns text directly from the source document rather than generating potentially fabricated quotations. - Evidence citations: Identifies supporting passages for decisions. - Calibrated probabilities: Provides confidence estimates designed to support reliable automation. THX-01 also supports TypeSafe-style choice , noul , and score questions, allowing integration with existing decision pipelines. How Does It Compare to Large Language Models? On our internal four-language support-ticket classification benchmark, THX-01 achieved 98.4% accuracy , compared with 98.5% for Claude Sonnet 5.5 . <escape That’s a 0.1 percentage-point difference, with approximately 150× faster inference in our tests.</escape Because THX-01 produces calibrated confidence estimates, our selective automation evaluation also showed that it could process approximately 75% of tickets automatically without errors on the evaluated subset . These are task-specific internal results, not claims of general equivalence to a large language model. THX-01 is optimized for structured decisions rather than open-ended generation. Built for Local AI One of our goals was to make useful AI decision-making accessible without massive infrastructure. With just 322 million parameters, THX-01 can run on consumer hardware, including phones. This makes it interesting for on-device classification, document processing, agent routing, and privacy-sensitive applications where sending every request to a large hosted model is unnecessary. Fully Open Source We’re releasing THX-01 under the Apache 2.0 license , including its model weights and inference tooling. You can install the Python package with a single command: pip install thx01 Run the model locally without an API key. Model and weights: https://huggingface.co/doofz/THX-01 https://huggingface.co/doofz/THX-01 What’s Next? We believe the future of AI infrastructure won’t be built entirely around increasingly large generative models. It will also require small, specialized models that make reliable decisions quickly, efficiently, and with measurable uncertainty. Not every AI task needs a generated answer. Sometimes, it just needs the right decision. THX-01 is our contribution to that direction. We’re excited to see what the open-source community builds with it.