SK Telecom (SKT) took the next step in a process of developing its own AI large language models (LLMs) by releasing a lightweight smartphone-focused version to expert open source community Hugging Face.

AX 3.1 Lite comprises 7 billion parameters and is the next version of a model employed for SKT’s A-dot call management AI feature.

The operator stated it developed the LLM from scratch, building on the performance and efficiency of the AX 3.0 Lite used in A-dot. SKT predicted the features of its latest model would deliver smooth services by enabling companies to deploy optimised set-ups on mobile devices.

SKT suggested information processing capacity and power consumption as areas the LLM could help to optimise.

AX 3.1 Lite “has excellent Korean language processing capabilities on par with its sister LLM AX 4.0 Lite”, SKT stated.

SKT previously revealed AX 4.0 Lite employs 72 billion parameters.

The operator stated AX 3.1 Lite achieved a 96 per cent performance level during Korean language proficiency assessment benchmark KMMLU2 and 102 per cent on CLIcK3, which adds culture to the mix.

AX 4.0 Lite achieved figures of 64.2 per cent and 71.2 per cent on the respective tests.

SKT noted it commenced work on its own LLMs in 2018, resulting in a leading position in the localisation and release of GPT2. It plans to make its “from scratch” method foundational in ongoing efforts in the field, which include advancing its sovereign AX 3 range along with using its AX 4 for continual pre-training “in areas that require optimised performance and efficiency”.

Hugging Face offers various collaborative assets to test AI functions.