iPhone 17 Pro Runs 400B Parameter Model Locally at 0.6 Tokens Per Second

What defines the computational ceiling of modern smartphones? Apple’s newest flagship, the iPhone17Pro , has provided an answer that is as impressive as it is humbling.
On March 23, a large language model featuring 400 billion parameters was successfully executed on the iPhone17Pro . Notably, even after quantization, such models typically demand at least 200GB of memory, whereas the iPhone17Pro is equipped with only 12GB of LPDDR5X RAM.
Technical Innovation: Flash Memory Streaming and Mixture of Experts (MoE)
Overcoming severe memory constraints, this "impossible feat" was primarily achieved through two key technical strategies:
SSD as Virtual Memory: Leveraging the open-source Flash-MoE project, the device streams data directly from the solid-state drive (SSD) to the GPU, effectively bypassing physical memory limitations.
MoE Architecture Benefits: "MoE" stands for Mixture of Experts, a design where the system activates only a subset of the 400 billion parameters for each generated word, rather than loading the entire model.
Performance Trade-off: Slow Generation Speed
While the execution was successful, the user experience remains far from practical. Test results indicate:
Generation Rate: The speed is limited to 0.6 Tokens per second, meaning it takes approximately 1.5 to 2 seconds to produce a single word.
Power and Thermal Impact: This intensive local processing rapidly depletes battery life and generates significant heat.
Industry Perspective: Is Local AI Reaching a "Singularity"?
Despite the current generation speed being suboptimal, the symbolic importance of this demonstration outweighs its immediate utility. It proves that running large-scale models locally on smartphones is a viable path.
Enhanced Privacy: Local processing ensures data remains on the device, eliminating the need for cloud uploads and offering superior privacy.
Offline Capability: Accessing advanced AI responses without an internet connection is becoming increasingly feasible.
Related article
Google updates Gemini app to compete with ChatGPT and Claude at IO 2026
During Tuesday’s annual Google I/O event, Google unveiled a suite of updates for the Gemini app, introducing a “Daily Brief” feature, a revamped interface, access to the new Gemini Omni AI video model, and a personal AI agent named Gemini Spark.These
Qianwen Creation Unveils 7-Day Membership With 200 Points and Fast-Track Tools for Travel Photography
Qwen Create has officially launched a 7-day membership plan designed for immediate, lightweight AI creation tasks, including travel photography, event marketing, and social media content. Upon purchase, users receive 200 points, access to exclusive f
iOS 27.2 Beta Brings Major Health App Overhaul and Siri Expands to Five Languages
Apple has just released the first iOS 27.2 beta for developers. Skipping the expected iOS 27.1 test phase, Apple launched the 27.2 testing cycle directly on September 16, pushing initial betas for iOS, iPadOS, macOS, watchOS, tvOS, and visionOS simul
Related Special Topic Recommendations
Comments (0)
0/500

What defines the computational ceiling of modern smartphones? Apple’s newest flagship, the
On March 23, a large language model featuring 400 billion parameters was successfully executed on the
Technical Innovation: Flash Memory Streaming and Mixture of Experts (MoE)
Overcoming severe memory constraints, this "impossible feat" was primarily achieved through two key technical strategies:
SSD as Virtual Memory: Leveraging the open-source Flash-MoE project, the device streams data directly from the solid-state drive (SSD) to the GPU, effectively bypassing physical memory limitations.
MoE Architecture Benefits: "MoE" stands for Mixture of Experts, a design where the system activates only a subset of the 400 billion parameters for each generated word, rather than loading the entire model.
Performance Trade-off: Slow Generation Speed
While the execution was successful, the user experience remains far from practical. Test results indicate:
Generation Rate: The speed is limited to 0.6 Tokens per second, meaning it takes approximately 1.5 to 2 seconds to produce a single word.
Power and Thermal Impact: This intensive local processing rapidly depletes battery life and generates significant heat.
Industry Perspective: Is Local AI Reaching a "Singularity"?
Despite the current generation speed being suboptimal, the symbolic importance of this demonstration outweighs its immediate utility. It proves that running large-scale models locally on smartphones is a viable path.
Enhanced Privacy: Local processing ensures data remains on the device, eliminating the need for cloud uploads and offering superior privacy.
Offline Capability: Accessing advanced AI responses without an internet connection is becoming increasingly feasible.
Google updates Gemini app to compete with ChatGPT and Claude at IO 2026
During Tuesday’s annual Google I/O event, Google unveiled a suite of updates for the Gemini app, introducing a “Daily Brief” feature, a revamped interface, access to the new Gemini Omni AI video model, and a personal AI agent named Gemini Spark.These
Qianwen Creation Unveils 7-Day Membership With 200 Points and Fast-Track Tools for Travel Photography
Qwen Create has officially launched a 7-day membership plan designed for immediate, lightweight AI creation tasks, including travel photography, event marketing, and social media content. Upon purchase, users receive 200 points, access to exclusive f
iOS 27.2 Beta Brings Major Health App Overhaul and Siri Expands to Five Languages
Apple has just released the first iOS 27.2 beta for developers. Skipping the expected iOS 27.1 test phase, Apple launched the 27.2 testing cycle directly on September 16, pushing initial betas for iOS, iPadOS, macOS, watchOS, tvOS, and visionOS simul





Home






