Four days ago we released THX-02, an open 4.5B decision model. You send a state (text, JSON or images) and typed questions; it returns a calibrated probability for every option you give it, and it writes a short thought first only when a question needs reasoning.

In those four days, more than 100,000,000 tokens went through THX-02 on our API. THX-02 also reached #1 in Zero-Shot Classification on Hugging Face. Here is what people built, with no names, just the shapes of the problems.

Developer tools and AI agents

The biggest group by far.

  • Pull-request review in CI. For every diff: does it expose a secret? For the PR description: does it claim something the code doesn't do, include unrelated changes, or contradict its tests?
  • Guardrails for coding agents. Does this command violate the policy, and did the user actually ask for it? Is a "done" claim backed by evidence? Is this session's learning worth saving to memory?
  • Routing. Which tool to call first, which use case a request belongs to, when to hand off to a human.

Decision models fit here because the answer must be one of a fixed set, fast, and come with a confidence you can threshold.

Documents, law and the public sector

  • Italian fact-checking: who speaks a quote, and whether a claim is supported by the passage.
  • Italian long-document Q&A: does any line answer this rule, and where.
  • Italian land and permits: landscape constraints and planning rules for solar.
  • French public procurement: would a company with this profile bid on this tender.
  • Indian GST bookkeeping: which ledger, asset or expense, and whether input tax credit is blocked.
  • Job postings: 15 fields per posting, from region and seniority to visa sponsorship.

Marketing, support and vision

  • Landing-page copy and UX screen audits, mostly with thinking on.
  • Support ticket triage in English, Chinese and Azerbaijani.
  • Receipts, tables and charts; video frames for sponsored-product visibility.

And the unexpected

An AI techno DJ asking which color, melodic and rhythm layer to bring in for the next 16 bars. A game studio choosing which recorded voice line replaces the hero's. An agent walking through a maze.

What we learned

  1. Decision models shine as checkers inside other systems: review, guardrails, moderation, routing. That's where confidence was highest (0.93 to 0.98 on average).
  2. People bring long, real inputs: diffs, legal documents, transcripts. Most tokens came from texts of several thousand tokens.
  3. They are honest about what they can't do. Numeric estimation (calories) and spatial puzzles got low confidence, which is the right behavior for a calibrated model, and a clear target for the next version.

Try it 

pip install thx02 thx02 "My order arrived broken, please send a new one" \ --choice "action=refund:money back,replace:send a new item,ignore:no action" \ --yesno "Does this ticket need a reply?"

We believe AI should be built for everyone and with everyone. Thank you to everyone who integrated THX-02 in its first week.

HAL-X, Baku