Apple has unveiled AFM 3 Core Advanced, a member of the third-generation Apple Foundation Models (AFM 3) developed in collaboration with Google. The model is designed to generate text and speech on certain Apple devices, optimized for Apple silicon to reduce active memory use compared with typical mixture-of-experts (MoE) models.
Capabilities and deployment
- Inputs/outputs: text, images and speech inputs; text and speech outputs.
- Architecture: a modified mixture-of-experts transformer with 20 billion parameters total and between 1 and 4 billion parameters active at any given time.
- Features: text understanding and generation, speech understanding and generation, image understanding, tool use, skills, and reasoning across 25 languages.
- Availability: slated to arrive in fall 2026 via Apple operating system updates for Macs and iPhone 17 Pro/Max/Air devices.
How it differs from standard MoE designs
AFM 3 Core Advanced uses an approach Apple calls Instruction-Following Pruning instead of the internal routing layers typical of MoE architectures. Conventional MoE models route each output token to particular experts via in-model routing layers, often switching experts token by token. AFM 3 Core Advanced employs a separate transformer to select which experts to activate for some or all output tokens; because the same experts are used across multiple tokens, the model performs fewer expert switches. This design can enable faster inference and lower memory usage than a typical MoE of similar size.
A practical benefit of this behavior is that the model becomes feasible to store on flash memory rather than needing all experts loaded into active RAM/VRAM at once, making a larger model runnable locally on the device.
Training data and transparency
Apple says AFM 3 models were trained on a mixture of publicly available information, licensed data, data collected from studies, and generated data. The company states that no user data or user interactions with the models were included in the training. The training process included pretraining, supervised fine-tuning, and reinforcement learning.
Apple has not disclosed input/output size limits, benchmark performance numbers, or the specific datasets and methods used in training. The company says it will release benchmark results later in the year. Apple also notes that AFM 3 models were custom-built and distilled from unspecified Google Gemini models.
Performance reporting
To date, Apple has not published independent benchmark results for AFM 3. The company says it will share results later in the year. Internally, Apple reports that AFM 3 models outperform the previous generation in proprietary human-preference measurements, but external comparative data are not yet available.
Context: Google partnership and developer choices
In January, Apple signed a multi-year agreement with Google to use Gemini models as the basis for its AI models. Amar Subramanya, Apple Vice President of AI, said at the AFM 3 launch that the models are "distillation-based, not a wholesale adoption of Gemini." Apple also announced the Foundation Models Framework, which will allow models from other companies to be used; developers working on Apple hardware will be able to choose between AFM 3 models and alternatives that implement Apple’s LanguageModel protocol, such as Anthropic Claude or Google Gemini families.
Why this matters
Downloading and running multi-billion-parameter models locally is typically impractical for many apps because of memory and bandwidth constraints. AFM 3 Core Advanced’s architecture addresses that limitation by reducing the need to load all experts into active memory, making it more practical to store and run a larger model in flash on-device. That creates a clear incentive for iOS developers to adopt streamlined, hardware-optimized models, and it leverages Apple’s partnership with Google to access expertise in running such models on phones.
Summary
AFM 3 Core Advanced is an Apple silicon–optimized, MoE-inspired model that contains 20 billion parameters but activates only 1–4 billion at a time. It will be distributed in fall 2026 with Apple OS updates for Macs and iPhone 17 Pro/Max/Air. Apple plans to publish benchmark and other technical details later in the year.
