Qualcomm CEO: Smartphones Running 100B LLMs Locally Could Arrive by 2028

Cristiano Amon said handsets will be able to run 100-billion-parameter language models continuously by 2028, and AI companies are already building such devices. He named no partners and did not address the DRAM supply and cost bottleneck.

• A 4-bit quantized 100B model needs at least 50 GB of RAM
• A 2-bit version still requires about 25 GB just for the model
• Workarounds include streaming from flash and MoE with expert offloading
• Distilled 20–30B models approach 100B quality on phones