On-device artificial intelligence (AI) will be a game-changer for tomorrow’s mobile industry across all enterprise applications, with AI being the crucial element in establishing new user paradigms and experiences, says Qualcomm Europe senior vice-president and president Enrico Salvatori.
Opening up at MWC 2025 about his company’s broad offer of new modems set for 5G networks and internet of things (IoT) applications, in addition to fixed wired access scenarios and general mobile infrastructure, the executive stressed just how quickly AI was scaling at the network edge to drive a number of key benefits. These include improvements in cost, performance and efficiency, privacy and security, immediacy, reliability, personalisation, and user interfaces and experiences.
In essence, said Salvatori, the mobile industry is moving into the era of a hybrid AI model, whereby model training is carried out using centralised cloud resources and the inference is done on the device using processors, principally an integrated Hexagon neural processor unit (NPU), Oryon central processing unit (CPU) and Adreno graphics processing unit (GPU). This, he said, will provide the cornerstone of edge AI leadership.
“Our vision [concerns] the impact on the device [and] the fact that the AI is introducing a new UI [user interface]. AI is moving from, or complemented by, the traditional cloud AI, also with the AI at the edge… We can provide much more NPU power to the device overall architecture … connected with fibre, Wi-Fi, etc. [More inference done on the device] offers benefits, first of all, because privacy and security of data are in the device.”
In terms of specific MWC announcements, Qualcomm unveiled its X85 5G Modem-RF eighth-generation 5G modem-to-antenna solution and fourth-generation AI-powered 5G connectivity system. The X85 is designed for the next generation of connected and AI-enabled applications, delivering faster speeds for seamless streaming, downloads and uploads, and is claimed to offer improved network reliability in congested areas, extended battery life and enhanced location accuracy for an overall superior user experience.
Shipping on Android smartphones, it is also engineered to support 5G Advanced capabilities in a multitude of other devices, including PCs, fixed wireless access points, automotive and XR headsets.
The X85 5G Modem-RF also has AI integrated, offering a number of previously unavailable performance benefits, said Salvatori. “The AI processor is optimising performance in terms of speed, in terms of coverage, and that is contributing coverage in terms of the environment. It is giving performance that you cannot achieve with a traditional CPU.”
Another key announcement was that the Dragonwing Cellular Infrastructure platforms for 5G Open RAN (O-RAN) solutions have been adopted by operators NTT Docomo and Viettel. The latter is Vietnam’s largest mobile network operator, which has deployed a live 5G O-RAN Massive MIMO network using Dragonwing Cellular Infrastructure platforms. Docomo is using Dragonwing X100 Accelerator Cards for its nationwide 5G vRAN network deployment across Japan.
The Dragonwing Fixed Wireless Access (FWA) Gen 4 Elite Platform, powered by the X85 5G Modem-RF, is said to be the world’s first 5G Advanced FWA platform, delivering what Qualcomm confidently expects to be “ultra-fast and most advanced” wireless mobile broadband experiences with integrated AI capabilities enhancing network performance and usher in an era of “unprecedented” generative AI innovation at the network edge.
Available immediately, the FWA platform design features a quad-core processor, dedicated hardware acceleration, integrated 5G Modem-RF, GNSS, Tri-band Wi-Fi 7 and broad carrier middleware support.
Qualcomm also announced that it was introducing three modem products to address the industrial internet of things (IIoT) – the E41, E51 and E52. These have 4G capability for entry-level performance, but with integrated RF, and the E51 and E52 have integrated GPS to boost modem connectivity.
As he set out the AI vision, Salvatori noted that AI models are evolving to small language models (SLMs) working with significantly reduced numbers of parameters. In a stark example of this evolution, he showed test results showing how AI capabilities on-device today equal that of last year’s on-cloud. He quoted research showing that the index level of model quality in March 2023 of the GPT 3.5 Turbo 175 billion parameter large language model was equalled in July 2024 by Llama 3.1 (SL) with eight billion parameters. Since then, other SLMs have produced similar, if not better, quality indices.
“There are more exports on the software side in designing new SLMs that can leverage increased hardware capacity into the device. It’s a kind of positive virtual circle, because we can optimise software so the machine is working in the right direction from our perspective. The device contains our data, our apps, and we can 1741066610 do more and create this new UI and user experience.
“An AI agent creates an overlay for the user, interfacing all apps, all the data available in the device, and next, through the language model multi model information. It can extract data from the internet, from the network, and all together, delivering the richer user experience. We no longer have to select which app, data file, or video to use – everything will be done by the AI agent.”
Another element the executive stressed was that the technology for these new experiences is already commercially available for an array of PC and smartphone partners, as well as a growing developer ecosystem. Driving this, he said, will be the Qualcomm AI Hub, which comprises a device farm, marketplace and centre for services, tools and models.
