AFM 3 and ADM model family architecture evolution
Parent: Apple Foundation Models framework · Published reference · snapshot 2026-10-23 · viability 4.20 / 5 (research)
↓ Facts as markdownall context files
Source-anchored research for AFM 3 and ADM model family architecture evolution. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Structure and components
- Google's January 12, 2026 statement says the next generation of Apple Foundation Models would be based on Gemini models and cloud technology. Apple's June announcement calls the five AFM 3 models custom-built in collaboration with Google, and NVIDIA describes the jointly built models as leveraging technologies behind Gemini. These official statements establish collaboration and Gemini-related technology use but do not identify exact checkpoint or weight transfers or specify which internal components were shared. (medium confidence) [source]
Measurements and reference values
- Apple's 2025 report describes two multilingual, multimodal language models: an approximately 3B on-device model with two-block KV-cache sharing and 2-bit quantization-aware training, and a server PT-MoE model using track parallelism and interleaved global/local attention. Apple reports the KV-cache design reduces cache memory by 37.5%. (high confidence) (verified as of 2025-08-27) [source]
Comparisons and alternatives
- Apple says AFM 3 Core Advanced uses Instruction-Following Pruning: prompt-time routing selects a fixed subset of experts from flash for loading into DRAM, combines routed experts with always-active shared experts, and periodically reselects experts during generation. The 20-billion-parameter model activates 1–4 billion parameters depending on the request. Apple's 2025 IFP paper describes the underlying method as an instruction-conditioned dynamic sparsity mask, but evaluates a distinct 3-billion-activated model rather than AFM 3 Core Advanced. (high confidence) [source]
- Apple reports quality evaluations at both the model and feature levels in 2026, and specifies that in-house human graders assess model-level responses. For general text preference, Apple reports AFM 3 Core at 45.6% versus 23.3% for its 2025 baseline, and AFM 3 Cloud at 64.7% versus 8.7% for its 2025 baseline. (low confidence) (verified as of 2026-06-08) [source]
- Apple's 2026 training description says the models shared an initial foundation before specializing, and that post-training combines supervised fine-tuning with multi-stage reinforcement learning. In Apple's in-house side-by-side human evaluations, AFM 3 Core was preferred on 45.6% of general-text prompts versus 23.3% for its 2025 baseline, while AFM 3 Cloud was preferred on 64.7% versus 8.7% for the 2025 server baseline. (medium confidence) (verified as of 2026-06-08) [source]
Facts and statements
- The 2024 technical report defines the AFM pair as an approximately 3-billion-parameter on-device language model and a larger server language model. It describes the AFM base models as dense decoder-only Transformers and places coding and diffusion/image models in the broader generative family. (high confidence) (verified as of 2026-05-27) [source]
- Apple's 2024 report says the wider generative-model family already included coding and diffusion models; its June 2026 article names ADM 3 Cloud (Image) for image generation and editing, so 2026 was not the first public disclosure of an Apple Intelligence image-model branch. (high confidence) (verified as of 2026-06-08) [source]
- In its June 8, 2026 announcement, Apple names five models: AFM 3 Core, a dense approximately 3-billion-parameter on-device model; AFM 3 Core Advanced, a natively multimodal 20-billion-parameter sparse on-device model with 1–4 billion active parameters; AFM 3 Cloud, a server workhorse; ADM 3 Cloud (Image), for image generation and editing; and AFM 3 Cloud Pro, for complex reasoning and agentic tool use. (high confidence) (verified as of 2026-06-08) [source]
- Apple describes AFM 3 Core as a next-generation approximately 3B dense model and AFM 3 Core Advanced as a natively multimodal 20B sparse model that activates 1–4B parameters per request. The disclosure adds a larger sparse on-device tier alongside the compact dense tier. (high confidence) (verified as of 2026-06-08) [source]
- Apple's 2026 training disclosure says the family shares a common initial foundation before models specialize for their architectures and use cases, followed by supervised fine-tuning and multi-stage reinforcement learning. (low confidence) [source]
- Apple's October 22, 2026 support article says next-generation Apple Intelligence is available with OS 27 releases on eligible devices and labels Siri AI as beta. The earlier June research article described its evaluation results as reflecting models at their current development stage and said improvements would continue during beta; neither page maps availability to each named AFM or ADM variant. (medium confidence) (verified as of 2026-10-22) [source]
- Apple's 2025 technical report describes adding image data and vision encoders to pretraining, then using supervised fine-tuning and reinforcement learning for post-training. Apple reports that the on-device and server models matched or exceeded comparable open baselines in its public benchmark and human evaluations. (medium confidence) [source]
- Apple says AFM 3 Cloud upgrades the PT-MoE foundation introduced in 2025 to stabilize training and improve reasoning and recall within the context window. Apple describes ADM 3 Cloud (Image) as a native image creation, editing, and Genmoji model that generalizes across aspect ratios and resolutions, with specialized adapters for downstream experiences such as Spatial Reframing and Image Playground modifications. (high confidence) [source]
Children
- No children recorded.