THX-01 averages 98.4 on its 2,843 support tickets

doofz posted THX-01 on October 11, 2026, at 10:23 UTC. The text says the model has 322 million parameters, beats Jev on some benchmarks, runs locally on a phone, and now has a free API. It links doofz/THX-01, the docs at api.hal-x.ai, and pip install thx01. The share image is the still on that post. An earlier post from the same account named the model before the card was opened.

The card says THX-01 is a non-autoregressive decision model from HAL-X AI. A state and one or more typed questions come back in one forward pass, about 10 ms on one GPU. The parameter line is 322M, with 307M in the encoder and 15M in the decision head. The encoder is mmBERT-base, 22 layers, a 256k vocabulary. Input is capped at 1,024 tokens per question, and longer states are truncated. License is Apache 2.0. The citation names Farid Aghayev and Elturan Ahmadbayli. Post-training is described as eight RLCD stages and about 2.1 million decisions, with Azerbaijani the largest language in the synthetic data. The card lists 18 post-training languages and says the encoder covers more than 100.

Question types on the card are choice, noul, and score, plus number, excerpt, and a cite flag. The card says number, excerpt, and citations are not TypeSafe question types. Number returns a value written in the document, or null when it is not stated, and the card says it does not convert units or currencies. Excerpt is cut from the document with character offsets.

The support-ticket table is fifteen categories and four sets, 2,843 tickets in Azerbaijani, Russian, English, and Turkish. Independent is described as written by GPT-6-Luna and kept out of training. Claude Sonnet 5.5 is marked as a stratified subset of 200 tickets per set.

Model	Clean	Corrupted	Messy	Independent	Average	ECE	Latency
THX-01	99.2	97.7	98.8	97.9	98.4	0.003	about 10 ms
Claude Sonnet 5.5	98.5	98.0	98.5	99.0	98.5		1.5 s
Wahoo 1.5	99.8	96.3	98.0	97.2	97.8		145 ms
GPT-6-Luna	98.6	97.4	96.7	98.2	97.7		1.9 s
Jev 1.13	99.2	95.0	96.2	99.0	97.4	0.007	331 ms
Kev-4B	96.1	88.4	90.8	95.8	92.8	0.202	830 ms

The card says ECE is measured on the Independent set, and that the LLMs return no probabilities. THX-01’s latency is on one GPU. The other clocks are through APIs. A later sentence says the model handles about three quarters of the tickets with no error on any of the four sets, and 90% of tickets at an accuracy of at least 99.5%. Those coverage lines are the card’s operating-point claim, not a second accuracy table.

Extraction is a different set: 1,824 number questions and 1,357 excerpt questions. Number lookup moves from 69.6 on the previous checkpoint to 93.4. The card says Jev 1.13 reaches 96.6 on number lookup over the same documents. Excerpt F1 is 84.1. Citation accuracy is 94.0. The 39-task suite mean is 84.2, with mean ECE 0.066. Banking77 inside that suite is 65.9, ECE 0.087. DAIR Emotion is 44.6. That Banking77 cell is not Cloudflare’s macro-F1 of 94.8 and not Bespoke Nimble’s 292 of 324.

The docs page documents POST https://api.hal-x.ai/v1/systemone, alias POST /v1/decide, model thx-01. A sample response on that page shows output_tokens at 0, a billing choice at probability 0.5239 beside refund at 0.4735, and latency_ms 22.94. The page does not print a price. The October 11 post calls the API free. Context on the docs page matches the card’s 1,024 tokens per question. Batch size is 1,024 items. Questions with more than 24 options use a two-round tournament. We did not call the API, and we did not run the weights.