Small and On-Device Models: Phi, Gemma, Apple — 2026-09-04

Top developments

Nvidia RTX Spark "Superchip" Debuts at IFA 2026

Nvidia and its partners showcased the first laptops and mini PCs powered by the new RTX Spark "Superchip" at IFA 2026. These devices are specifically engineered to run AI models directly on the computer, reducing reliance on cloud APIs for local inference tasks. This hardware launch complements the recent trend of small language models like Microsoft's Phi-4-mini and Google's Gemma 3 becoming viable on consumer-grade NPUs and GPUs.

Samsung Galaxy S26 FE Launches with Exynos 2500 and On-Device AI

Samsung officially launched the Galaxy S26 FE, featuring One UI 9 and a suite of new AI-powered tools including Photo Assist and Horizon Lock. The device is expected to utilize the Exynos 2500 chipset, signaling Samsung's continued commitment to optimizing on-device AI performance for its mid-range flagship line. This move aligns with broader industry efforts to integrate efficient small language models directly into smartphone hardware for faster, privacy-focused processing.

South Korea Finalizes K-On Device AI Semiconductor Consortium

After months of delays due to budget disputes between ministries, South Korea has finalized the final partners for its national "K-On Device AI Semiconductor" project. The consortium, which includes notable players like Mobilit and HyperExcel, aims to develop specialized chips for edge AI applications. This government-backed initiative underscores the strategic importance of localizing AI hardware production to support domestic on-device AI ecosystems.

Local view

South Korean media outlets have closely followed the finalization of the K-On Device AI project, with ZDNet Korea highlighting the selection of key consortium members after a period of uncertainty. Additionally, user feedback on Samsung’s community forums indicates growing scrutiny of UI clarity regarding "on-device only" data processing settings, reflecting heightened consumer awareness of privacy features in on-device AI implementations.

Context & numbers

The push for on-device AI is supported by competitive performance metrics from small language models. Recent benchmarks indicate that Microsoft’s Phi-4-mini (3.8B parameters) achieves an 83.7% score on ARC-C, a top result in its size class, while Google’s Gemma 3 4B posts an 89.2% on GSM8K math reasoning. These figures demonstrate that sub-10B models are increasingly capable of handling complex tasks locally, driving demand for hardware like the newly announced Nvidia RTX Spark.

On the radar

• DeepX NPU Expansion: South Korean startup DeepX reported 77 global orders in its first year of mass production for its ultra-low-power DX-M1 NPU, indicating strong early adoption for edge AI devices.
• Hugging Face Microduck Robot: The small robot "Microduck," co-developed by Hugging Face and Pollen Robotics, has seen explosive initial sales, featuring a Chinese Rockchip NPU, highlighting the integration of small models into consumer robotics.