Apple has introduced its third-generation foundation models, AFM 3, which allow for a 20-billion-parameter on-device AI model by storing weights in NAND flash instead of DRAM, thus overcoming previous memory limitations. This new architecture enables enterprises to run complex AI tasks locally, while also providing server-based options for more demanding workloads, although details on deployment and performance metrics remain pending.
The most valuable insight for you is the innovative approach Apple has taken with their AFM 3 Core Advanced model by storing its entire 20-billion-parameter set in NAND flash rather than DRAM. This architectural shift allows for more complex on-device AI models without relying on cloud infrastructure, a significant development for enterprises requiring high-capacity AI models that operate independently of a continuous cloud connection. This could influence your strategies for deploying AI workloads, enabling more robust on-device capabilities.