LG Uplus says two research papers on smartphone AI have been accepted at NeurIPS 2026’s On-Device Intelligence workshop. One reports a model running about 4.5 times faster while using about 77 per cent less battery than existing methods, a potentially useful step for AI that does not have to live in the cloud.
Watch Desk analysis
What happened
The South Korean telecoms company says it developed a small language model for call services and optimised it for smartphones with limited memory and power. LG Uplus attributes the speed and battery figures to that work, and says the model performed on a par with the latest cloud-based AI.
A second paper describes training AI on a smartphone without sending personal data outside the device. LG Uplus says the method achieved accuracy on a par with server-based training. The company plans to apply the technology to its AI services and expand the research to servers, according to The Korea Times.
Why it matters
On-device AI can mean quicker responses and less reliance on a remote service. Keeping personal data on the phone during training, if the reported method holds up in practice, could also give developers another way to limit data leaving a device.
The headline figures are promising, but the available account does not give benchmark details or enough information to compare the methods independently. That makes the papers and the results behind them the next things to examine, rather than treating the numbers as a settled verdict.
Our read
This is a useful signal that phone-based AI research is tackling two practical constraints at once: battery life and data leaving the device. LG Uplus says it intends to bring the work into services, which gives the research a path beyond the conference programme. The real test will be whether the gains survive real-world use, where phones have a habit of being less cooperative than a benchmark.
What to watch
- What the accepted papers disclose about their methods and comparisons.
- Whether LG Uplus demonstrates the model in a real service.
- Whether the on-device training approach works beyond the company’s reported results. Activity teaser LG Uplus says its smartphone AI model ran about 4.5 times faster and used about 77 per cent less battery than existing methods, while performing on a par with the latest cloud-based AI. The company also reports a way to train AI on a phone without sending personal data elsewhere. Both papers have been accepted for NeurIPS 2026’s On-Device Intelligence workshop. The figures come from LG Uplus, and the detail behind the comparisons matters. Should phone AI prioritise keeping data on-device, or is cloud-level performance the more important prize?
Discussion spark: For AI on phones, which should matter more: keeping personal data on-device, or matching the performance of cloud-based systems?
Sources and evidence
- Source update (29 October 2026, 05:18 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.