THX-01
An open-source non-autoregressive decision model released by HAL-X AI, with 322 million parameters. It is trained using the RLCD method and supports six structured decision types including classification, yes/no judgment, and numerical extraction. Single-inference latency is approximately 10 milliseconds.
Timeline (last 90 days)
THX-01 is a non-autoregressive decision model open-sourced by HAL-X AI, with 322 million (322M) parameters.
Since its launch, THX-01's free API has processed more than 70 million tokens.
THX-01 returns structured answers with calibrated probabilities in a single forward pass; official GPU benchmarks show that a single inference takes approximately 10 milliseconds.
On an internal benchmark of 2,843 multilingual customer service tickets, THX-01 achieved 98.4% accuracy, an ECE of 0.003, and a measured latency of approximately 10ms, outperforming TypeSafe Jev 1.13 (97.4% accuracy, ECE 0.007, latency 331ms).
The latency figures for THX-01 and Jev 1.13 come from different service environments and are not a rigorous comparison conducted under identical hardware conditions.
THX-01 supports six decision types: Choice, Yes/No, Score, Number, Excerpt, and Citations.
THX-01 supports 18 post-training languages, including English, Russian, Turkish, and Azerbaijani, and can run locally on a smartphone.
THX-01's API provides a free quota of 200 decisions per minute.
THX-01 is open-sourced under the Apache 2.0 license; model weights, source code, and training resources are all publicly available on Hugging Face, and the Python package can be installed via pip install thx01.
All Facts (9)
THX-01 is a non-autoregressive decision model open-sourced by HAL-X AI, with 322 million (322M) parameters.
50%UnverifiedSince its launch, THX-01's free API has processed more than 70 million tokens.
50%UnverifiedTHX-01 returns structured answers with calibrated probabilities in a single forward pass; official GPU benchmarks show that a single inference takes approximately 10 milliseconds.
50%UnverifiedOn an internal benchmark of 2,843 multilingual customer service tickets, THX-01 achieved 98.4% accuracy, an ECE of 0.003, and a measured latency of approximately 10ms, outperforming TypeSafe Jev 1.13 (97.4% accuracy, ECE 0.007, latency 331ms).
50%UnverifiedThe latency figures for THX-01 and Jev 1.13 come from different service environments and are not a rigorous comparison conducted under identical hardware conditions.
50%UnverifiedTHX-01 supports six decision types: Choice, Yes/No, Score, Number, Excerpt, and Citations.
50%UnverifiedTHX-01 supports 18 post-training languages, including English, Russian, Turkish, and Azerbaijani, and can run locally on a smartphone.
50%UnverifiedTHX-01's API provides a free quota of 200 decisions per minute.
50%UnverifiedTHX-01 is open-sourced under the Apache 2.0 license; model weights, source code, and training resources are all publicly available on Hugging Face, and the Python package can be installed via pip install thx01.
50%