AI models are increasingly running directly on smartphones, completely without an internet connection or cloud servers. Apple integrates two of its own foundation models, AFM 3 Core and AFM 3 Core Advanced, directly into iOS, while Google offers comparable components for Android with Gemini Nano and the open Gemma 4. The performance of these models varies significantly in terms of parameter count, context window, and supported input formats.
Apple delivers two foundation models directly on the iPhone
Apple’s base model AFM 3 Core operates with around three billion parameters and runs on all Apple Intelligence-capable devices. It processes text and images as input and outputs text. The context window is shared between input and output: According to Apple’s developer documentation, a total of 4,096 tokens are available, of which about 96 tokens remain for the response. At WWDC on June 8, 2026, Apple introduced a second, significantly larger model, AFM 3 Core Advanced. It utilizes a sparse architecture with 20 billion parameters but activates only one to four billion of them per request. The model is, according to Apple Machine Learning Research, natively multimodal and additionally processes audio. Unlike the base model, AFM 3 Core Advanced runs only on the iPhone 17 Pro and on iPad and Mac models with A19 Pro, M3, or M4 chips; devices with eight gigabytes of RAM are excluded. For tasks that exceed the device’s capabilities, Apple’s Private Cloud Compute steps in as needed with a significantly larger context window of 32,000 tokens – but then no longer purely locally.
Google brings Gemini Nano and the open Gemma 4 to Android
On the Android side, Gemini Nano runs via the system service AICore on compatible devices like Pixel smartphones. Google has not yet published official values for parameter count and context window. Since April 2, 2026, Google has also offered Gemma 4 in the AICore developer preview – an open model in two sizes. E2B has, according to Google’s model card, 2.3 billion effective parameters (5.1 billion including embeddings), while E4B has 4.5 billion effective parameters (8 billion including embeddings). Both variants offer a context window of 128,000 tokens, process text, images, and audio, and support more than 140 languages. E2B is optimized for speed and is about three times faster than E4B, which, in turn, provides more computing power for more complex tasks. The models run on AICore-capable devices with neural processing units from Google, MediaTek, or Qualcomm.
Open models like Gemma 4 now also run on the iPhone
Gemma 4 is not an Apple product, but it is openly available – and can therefore also be run on the iPhone, for example via Google’s own LiteRT-LM, Apple’s framework MLX, or the open-source llama.cpp. A benchmark published in May 2026 measured about 37.8 tokens per second for Gemma 4 E2B on an iPhone 17 using llama.cpp, while Google’s own LiteRT-LM achieved around 55.4 tokens per second. Another example of open models on the iPhone is Bonsai 27B from PrismML: The California startup compresses a 27-billion-parameter model through quantization to 3.9 to 5.9 gigabytes while reportedly maintaining a context window of 262,000 tokens. Through the Locally AI app, it runs on iPhones from model 15 with around eleven tokens per second – these values come from PrismML itself and are independently unverified. Microsoft also offers open models with Phi-4-mini (3.8 billion parameters, pure text model) and Phi-4-multimodal (5.6 billion parameters, additionally processes image and speech) with a context window of 128,000 tokens each, which can be used locally on iPhone and Android via apps based on MLX or ONNX Runtime.
| Model | Platform | Effective Parameters | Context Window | Input | Output |
|---|---|---|---|---|---|
| Apple AFM 3 Core | iPhone (Apple Intelligence) | approx. 3 billion | 4,096 tokens (shared) | Text, Image | Text |
| Apple AFM 3 Core Advanced | iPhone 17 Pro, iPad/Mac with M3/M4 | 20 billion (1–4 billion active) | not specified by Apple | Text, Image, Audio | Text |
| Google Gemini Nano | Android (AICore) | not specified by Google | not specified by Google | Text | Text |
| Google Gemma 4 E2B | Android (AICore), iPhone (third-party) | 2.3 billion (5.1 billion with embeddings) | 128,000 tokens | Text, Image, Audio | Text |
| Google Gemma 4 E4B | Android (AICore), iPhone (third-party) | 4.5 billion (8 billion with embeddings) | 128,000 tokens | Text, Image, Audio | Text |
| Microsoft Phi-4-mini | iPhone/Android (third-party apps) | 3.8 billion | 128,000 tokens | Text | Text |
| Microsoft Phi-4-multimodal | iPhone/Android (third-party apps) | 5.6 billion | 128,000 tokens | Text, Image, Audio | Text |
| PrismML Bonsai 27B (compressed) | iPhone from model 15 | 27 billion (compressed to 3.9–5.9 GB) | 262,000 tokens (according to PrismML) | Text, Image | Text |
It will be crucial whether Apple opens AFM 3 Core Advanced beyond the currently exclusive iPhone 17 Pro and M3/M4 devices or whether open alternatives like Gemma 4 and Bonsai 27B continue to close the gap on older devices. For Google, it remains open when Gemini Nano will receive complete technical specifications or whether Gemma 4 will take on a leading role in AICore in the long term.
