Small and On-Device Models: Phi, Gemma, Apple — 2026-09-05

Top developments

Nvidia Unveils RTX Spark for Local AI at IFA 2026

At IFA 2026 in Berlin, Nvidia and its partners showcased the first laptops and mini PCs powered by the new "RTX Spark" superchip. These devices are explicitly designed to run AI models directly on the computer, marking a significant step in making high-performance local inference accessible to mainstream consumers. The launch highlights the industry's pivot from cloud-dependent AI to hybrid models that leverage powerful local NPUs and GPUs.

South Korea Finalizes "K-On Device" AI Semiconductor Consortium

South Korea’s government has finalized the consortium for its national "K-On Device" AI semiconductor project after months of budget-related delays. The project, which aims to develop domestic chips for edge AI applications, selected Mobilin and HyperExcel as key partners. This move is critical for reducing reliance on foreign chipmakers and fostering a local ecosystem for small language models and on-device AI applications.

Lenovo IdeaPad Vibe Launches with NPU Caveats

Lenovo announced the IdeaPad Vibe at IFA 2026, a laptop featuring AMD Ryzen AI 400 and Snapdragon X series chips. While both chipsets include NPUs eligible for Copilot+ features, the base model’s 8GB RAM configuration prevents it from qualifying for Copilot+ AI features, highlighting the ongoing tension between hardware capabilities and memory requirements for running local SLMs like Phi or Gemma. AMD models are scheduled to arrive in October 2026.

Samsung Community Pushes for Clearer On-Device AI Settings

Samsung Galaxy users are actively requesting UI improvements for the "Process data only on device" setting in Galaxy AI. Users report confusion over whether toggling the switch ON or OFF actually enforces local processing, indicating a need for better transparency in how Samsung’s on-device models handle privacy and data flow. This feedback loop suggests that user trust in on-device AI features remains a key challenge for vendors.

Local view

ZDNet Korea reports that the finalization of the K-On Device consortium is a pivotal moment for South Korea's tech sector, aiming to create a sovereign supply chain for edge AI chips. The involvement of firms like Mobilin and HyperExcel suggests a focus on specialized accelerators that can run quantized models efficiently, reducing latency and power consumption compared to general-purpose CPUs.

Samsung Members Community discussions reveal growing user awareness about data privacy in on-device AI. Users are scrutinizing the "on-device only" toggles in Galaxy AI, demanding clearer indicators of when data leaves the device. This reflects a broader trend where users are becoming more technically literate about the trade-offs between cloud-based and local AI processing.

Context & numbers

• Hardware Requirements: The launch of devices like the Lenovo IdeaPad Vibe underscores that 8GB of RAM is often insufficient for full Copilot+ or advanced on-device AI experiences, pushing the baseline higher for users wanting to run local SLMs effectively.
• Industry Shift: Nvidia's RTX Spark launch at IFA 2026 signals that major silicon vendors are now prioritizing consumer hardware capable of running substantial local models, moving beyond niche developer kits.

On the radar

• October 2026: Lenovo's AMD Ryzen AI 400-equipped IdeaPad Vibe models are scheduled for release, offering a test case for how AMD's latest NPUs handle local inference compared to Snapdragon and Apple Silicon.
• K-On Device Project Milestones: Watch for further announcements from the newly formed South Korean consortium regarding prototype chips and partnerships with mobile OEMs like Samsung.