Small and On-Device Models: Phi, Gemma, Apple — 2026-10-13
Qualcomm CEO Cristiano Ammon revealed that major AI firms are demanding chips capable of running 100-billion parameter models on smartphones by 2028, signaling a massive leap in on-device AI capabilities. Meanwhile, Microsoft launched new "Hybrid Intelligence" features for Copilot+ PCs, enabling local model execution and file-based actions, while Samsung showcased its latest AI-integrated devices at KES 2026.
Small and On-Device Models: Phi, Gemma, Apple — 2026-10-13
Qualcomm Targets 100B On-Device Models for Smartphones
During the Snapdragon Summit 2026, Qualcomm CEO Cristiano Ammon stated that leading AI companies are requesting chipsets capable of running models with at least 100 billion parameters directly on smartphones by 2028. This ambition far exceeds current on-device capabilities, which typically support models in the 1B–14B range (such as Phi-4-mini or Gemma 3 4B), requiring significant advancements in NPU efficiency and memory bandwidth to handle the increased computational load and power consumption.
Microsoft Unveils "Hybrid Intelligence" for Local Copilot Execution
Microsoft announced a new architecture called "Hybrid Intelligence" that allows Copilot to read local files and execute actions on the PC using locally hosted AI models rather than relying solely on cloud processing. This feature is restricted to Copilot+ PCs with specific NPU requirements, excluding devices with powerful discrete GPUs but lacking the designated NPU architecture, and will roll out starting October 16 with new Nvidia-powered Surface laptops.
Samsung Showcases AI-Integrated Devices at KES 2026
Samsung Electronics is participating in the Korea Electronics Show (KES) 2026, running from October 13–16 in Seoul, displaying its latest AI technologies across displays, home appliances, and mobile devices. The exhibition highlights how Samsung is integrating on-device AI into consumer hardware, leveraging its Exynos NPUs and Galaxy AI features to enhance user experience without constant cloud dependency.
Domestic Korean NPU Firms Demonstrate On-Device AI Hardware
At AI Festa 2026, South Korean semiconductor companies including Analog Plus, P&C Solution, and Rhinos displayed on-device AI devices powered by domestic Neural Processing Units (NPUs) from Telechips, DeepX, and Duffer Eye. This event underscores the growing momentum of the Korean domestic semiconductor industry in supporting lightweight, efficient AI inference for edge devices, reducing reliance on foreign chipsets for specific industrial and consumer applications.
Local view
Korean tech media are closely monitoring the shift toward domestic on-device AI infrastructure. ZDNet Korea highlighted the debut of local NPU-powered devices at AI Festa 2026, emphasizing the strategic importance of "K-Semiconductor" self-sufficiency in edge AI. The coverage suggests a national push to build a supply chain where domestic chips can run quantized small language models (SLMs) efficiently, aligning with government initiatives to localize critical AI infrastructure.
Additionally, Hankyung reported on how the global AI paradigm shift from training to inference presents a unique opportunity for Korean system semiconductor companies. By focusing on low-power, high-efficiency inference chips suitable for on-device SLMs, Korean firms aim to capture a larger share of the edge AI market, moving beyond traditional memory and logic roles.
Context & numbers
The demand for higher-capacity on-device models is driving rapid changes in hardware specifications. While current state-of-the-art on-device models like Gemma 4 E4B or Phi-4-mini operate within 2–8 GB of VRAM/RAM, Qualcomm’s projection of 100B-parameter models on phones implies a need for multi-gigabyte unified memory architectures and extreme quantization techniques (likely sub-4-bit) to fit such weights into mobile form factors.
Microsoft’s restriction of Hybrid Intelligence to Copilot+ PCs reinforces the industry standard that NPU performance (measured in TOPS) is now a gating factor for local AI features, not just raw GPU power. New devices starting at $2,599 (Surface Laptop) will ship with these capabilities, setting a price floor for full-featured local AI PCs.
On the radar
- October 16, 2026: Microsoft's new Nvidia-powered Surface laptops with Hybrid Intelligence begin shipping, offering the first real-world test of their local file-action capabilities.
- October 13–16, 2026: KES 2026 continues in Seoul, providing further details on Samsung’s specific on-device AI implementations in upcoming Galaxy and appliance products.
- Q4 2026: Rumors suggest Apple may announce updated Foundation Models for iOS 27/macOS 27, potentially increasing parameter counts for on-device tasks, though no official date has been confirmed.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
