Introducing Apple’s On-Machine and Server Basis Models

On the 2024 Worldwide Developers Conference, we launched Apple Intelligence, a private intelligence system built-in deeply into iOS 18, iPadOS 18, and macOS Sequoia.

Apple Intelligence is comprised of a number of highly-capable generative fashions which might be specialised for our customers’ on a regular basis duties, and may adapt on the fly for his or her present exercise. The inspiration fashions constructed into Apple Intelligence have been fine-tuned for person experiences comparable to writing and refining textual content, prioritizing and summarizing notifications, creating playful photos for conversations with household and associates, and taking in-app actions to simplify interactions throughout apps.

Within the following overview, we are going to element how two of those fashions — a ~3 billion parameter on-device language mannequin, and a bigger server-based language mannequin out there with Private Cloud Compute and working on Apple silicon servers — have been constructed and tailored to carry out specialised duties effectively, precisely, and responsibly. These two basis fashions are half of a bigger household of generative fashions created by Apple to assist customers and builders; this features a coding mannequin to construct intelligence into Xcode, in addition to a diffusion mannequin to assist customers specific themselves visually, for instance, within the Messages app. We sit up for sharing extra data quickly on this broader set of fashions.

Our Concentrate on Accountable AI Improvement

Apple Intelligence is designed with our core values at each step and constructed on a basis of groundbreaking privateness improvements.

Moreover, we’ve created a set of Accountable AI ideas to information how we develop AI instruments, in addition to the fashions that underpin them:

  1. Empower customers with clever instruments: We determine areas the place AI can be utilized responsibly to create instruments for addressing particular person wants. We respect how our customers select to make use of these instruments to perform their targets.
  2. Symbolize our customers: We construct deeply private merchandise with the aim of representing customers across the globe authentically. We work repeatedly to keep away from perpetuating stereotypes and systemic biases throughout our AI instruments and fashions.
  3. Design with care: We take precautions at each stage of our course of, together with design, mannequin coaching, function improvement, and high quality analysis to determine how our AI instruments could also be misused or result in potential hurt. We are going to repeatedly and proactively enhance our AI instruments with the assistance of person suggestions.
  4. Defend privateness: We defend our customers’ privateness with highly effective on-device processing and groundbreaking infrastructure like Personal Cloud Compute. We don’t use our customers’ non-public private knowledge or person interactions when coaching our basis fashions.

These ideas are mirrored all through the structure that permits Apple Intelligence, connects options and instruments with specialised fashions, and scans inputs and outputs to offer every function with the data wanted to perform responsibly.

Within the the rest of this overview, we offer particulars on choices comparable to: how we develop fashions which might be extremely succesful, quick, and power-efficient; how we strategy coaching these fashions; how our adapters are fine-tuned for particular person wants; and the way we consider mannequin efficiency for each helpfulness and unintended hurt.

Modeling overview
Determine 1: Modeling overview for the Apple basis fashions.


Our basis fashions are skilled on Apple’s AXLearn framework, an open-source venture we launched in 2023. It builds on prime of JAX and XLA, and permits us to coach the fashions with excessive effectivity and scalability on numerous coaching {hardware} and cloud platforms, together with TPUs and each cloud and on-premise GPUs. We used a mix of knowledge parallelism, tensor parallelism, sequence parallelism, and Absolutely Sharded Information Parallel (FSDP) to scale coaching alongside a number of dimensions comparable to knowledge, mannequin, and sequence size.

We practice our basis fashions on licensed knowledge, together with knowledge chosen to boost particular options, in addition to publicly out there knowledge collected by our web-crawler, AppleBot. Net publishers have the option to opt out of using their internet content material for Apple Intelligence coaching with a knowledge utilization management.

We by no means use our customers’ non-public private knowledge or person interactions when coaching our basis fashions, and we apply filters to take away personally identifiable data like social safety and bank card numbers which might be publicly out there on the Web. We additionally filter profanity and different low-quality content material to stop its inclusion within the coaching corpus. Along with filtering, we carry out knowledge extraction, deduplication, and the applying of a model-based classifier to determine top quality paperwork.


We discover that knowledge high quality is important to mannequin success, so we make the most of a hybrid knowledge technique in our coaching pipeline, incorporating each human-annotated and artificial knowledge, and conduct thorough knowledge curation and filtering procedures. Now we have developed two novel algorithms in post-training: (1) a rejection sampling fine-tuning algorithm with instructor committee, and (2) a reinforcement studying from human suggestions (RLHF) algorithm with mirror descent coverage optimization and a leave-one-out benefit estimator. We discover that these two algorithms result in vital enchancment within the mannequin’s instruction-following high quality.


Along with making certain our generative fashions are extremely succesful, we’ve used a spread of revolutionary methods to optimize them on-device and on our non-public cloud for velocity and effectivity. Now we have utilized an in depth set of optimizations for each first token and prolonged token inference efficiency.

Each the on-device and server fashions use grouped-query-attention. We use shared enter and output vocab embedding tables to cut back reminiscence necessities and inference price. These shared embedding tensors are mapped with out duplications. The on-device mannequin makes use of a vocab measurement of 49K, whereas the server mannequin makes use of a vocab measurement of 100K, which incorporates further language and technical tokens.

For on-device inference, we use low-bit palletization, a crucial optimization method that achieves the required reminiscence, energy, and efficiency necessities. To take care of mannequin high quality, we developed a brand new framework utilizing LoRA adapters that includes a combined 2-bit and 4-bit configuration technique — averaging 3.5 bits-per-weight — to realize the identical accuracy because the uncompressed fashions.

Moreover, we use an interactive mannequin latency and energy evaluation software, Talaria, to raised information the bit charge choice for every operation. We additionally make the most of activation quantization and embedding quantization, and have developed an strategy to allow environment friendly Key-Worth (KV) cache replace on our neural engines.

With this set of optimizations, on iPhone 15 Professional we’re capable of attain time-to-first-token latency of about 0.6 millisecond per immediate token, and a technology charge of 30 tokens per second. Notably, this efficiency is attained earlier than using token hypothesis methods, from which we see additional enhancement on the token technology charge.

Mannequin Adaptation

Our basis fashions are fine-tuned for customers’ on a regular basis actions, and may dynamically specialize themselves on-the-fly for the duty at hand. We make the most of adapters, small neural community modules that may be plugged into numerous layers of the pre-trained mannequin, to fine-tune our fashions for particular duties. For our fashions we adapt the eye matrices, the eye projection matrix, and the totally linked layers within the point-wise feedforward networks for an appropriate set of the decoding layers of the transformer structure.

By fine-tuning solely the adapter layers, the unique parameters of the bottom pre-trained mannequin stay unchanged, preserving the final information of the mannequin whereas tailoring the adapter layers to assist particular duties.

Determine 2: Adapters are small collections of mannequin weights which might be overlaid onto the widespread base basis mannequin. They are often dynamically loaded and swapped — giving the muse mannequin the power to specialize itself on-the-fly for the duty at hand. Apple Intelligence features a broad set of adapters, every fine-tuned for a particular function. It’s an environment friendly solution to scale the capabilities of our basis mannequin.

We characterize the values of the adapter parameters utilizing 16 bits, and for the ~3 billion parameter on-device mannequin, the parameters for a rank 16 adapter usually require 10s of megabytes. The adapter fashions could be dynamically loaded, briefly cached in reminiscence, and swapped — giving our basis mannequin the power to specialize itself on the fly for the duty at hand whereas effectively managing reminiscence and guaranteeing the working system’s responsiveness.

To facilitate the coaching of the adapters, we created an environment friendly infrastructure that permits us to quickly retrain, take a look at, and deploy adapters when both the bottom mannequin or the coaching knowledge will get up to date. The adapter parameters are initialized utilizing the accuracy-recovery adapter launched within the Optimization part.

Efficiency and Analysis

Our focus is on delivering generative fashions that may allow customers to speak, work, specific themselves, and get issues accomplished throughout their Apple merchandise. When benchmarking our fashions, we concentrate on human analysis as we discover that these outcomes are extremely correlated to person expertise in our merchandise. We carried out efficiency evaluations on each feature-specific adapters and the muse fashions.

As an example our strategy, we have a look at how we evaluated our adapter for summarization. As product necessities for summaries of emails and notifications differ in refined however necessary methods, we fine-tune accuracy-recovery low-rank (LoRA) adapters on prime of the palletized mannequin to fulfill these particular necessities. Our coaching knowledge relies on artificial summaries generated from greater server fashions, filtered by a rejection sampling technique that retains solely the top quality summaries.

To judge the product-specific summarization, we use a set of 750 responses rigorously sampled for every use case. These analysis datasets emphasize a various set of inputs that our product options are more likely to face in manufacturing, and embrace a stratified combination of single and stacked paperwork of various content material sorts and lengths. As product options, it was necessary to judge efficiency in opposition to datasets which might be consultant of actual use circumstances. We discover that our fashions with adapters generate higher summaries than a comparable mannequin.

As a part of accountable improvement, we recognized and evaluated particular dangers inherent to summarization. For instance, summaries often take away necessary nuance or different particulars in methods which might be undesirable. Nevertheless, we discovered that the summarization adapter didn’t amplify delicate content material in over 99% of focused adversarial examples. We proceed to adversarially probe to determine unknown harms and develop our evaluations to assist information additional enhancements.

Determine 3: Ratio of “good” and “poor” responses for 2 summarization use circumstances relative to all responses. Summaries are categorized as “good”, “impartial”, “poor” given the grader’s scores throughout 5 dimensions. A result’s categorized as “good” if the entire dimensions are good (increased is best). A result’s categorized as “poor” if any of the size are poor (decrease is best). Our fashions with adapters generate higher summaries than a comparable mannequin.

Along with evaluating function particular efficiency powered by basis fashions and adapters, we consider each the on-device and server-based fashions’ normal capabilities. We make the most of a complete analysis set of real-world prompts to check the final mannequin capabilities. These prompts are various throughout totally different issue ranges and canopy main classes comparable to brainstorming, classification, closed query answering, coding, extraction, mathematical reasoning, open query answering, rewriting, security, summarization, and writing.

We examine our fashions with each open-source fashions (Phi-3, Gemma, Mistral, DBRX) and industrial fashions of comparable measurement (GPT-3.5-Turbo, GPT-4-Turbo)1. We discover that our fashions are most popular by human graders over most comparable competitor fashions. On this benchmark, our on-device mannequin, with ~3B parameters, outperforms bigger fashions together with Phi-3-mini, Mistral-7B, and Gemma-7B. Our server mannequin compares favorably to DBRX-Instruct, Mixtral-8x22B, and GPT-3.5-Turbo whereas being extremely environment friendly.

Determine 4: Fraction of most popular responses in side-by-side analysis of Apple’s basis mannequin in opposition to comparable fashions. We discover that our fashions are most popular by human graders.

We use a set of various adversarial prompts to check the mannequin efficiency on dangerous content material, delicate matters, and factuality. We measure the violation charges of every mannequin as evaluated by human graders on this analysis set, with a decrease quantity being fascinating. Each the on-device and server fashions are strong when confronted with adversarial prompts, attaining violation charges decrease than open-source and industrial fashions.

Determine 5: Fraction of violating responses for dangerous content material, delicate matters, and factuality (decrease is best). Our fashions are strong when confronted with adversarial prompts.

Our fashions are most popular by human graders as protected and useful over competitor fashions for these prompts. Nevertheless, contemplating the broad capabilities of huge language fashions, we perceive the limitation of our security benchmark. We’re actively conducting each handbook and computerized red-teaming with inside and exterior groups to proceed evaluating our fashions’ security.

Determine 6: Fraction of most popular responses in side-by-side analysis of Apple’s basis mannequin in opposition to comparable fashions on security prompts. Human graders discovered our responses safer and extra useful.

To additional consider our fashions, we use the Instruction-Following Eval (IFEval) benchmark to check their instruction-following capabilities with fashions of comparable measurement. The outcomes recommend that each our on-device and server mannequin comply with detailed directions higher than the open-source and industrial fashions of comparable measurement.

Determine 7: Instruction-following functionality (measured with IFEval) for Apple’s basis fashions and fashions of comparable measurement (increased is best).

We consider our fashions’ writing capacity on our inside summarization and composition benchmarks, consisting of quite a lot of writing directions. These outcomes don’t confer with our feature-specific adapter for summarization (seen in Figure 3), nor do we’ve an adapter targeted on composition.

Determine 8: Writing capacity on inside summarization and composition benchmarks (increased is best).


The Apple basis fashions and adapters launched at WWDC24 underlie Apple Intelligence, the brand new private intelligence system that’s built-in deeply into iPhone, iPad, and Mac, and permits highly effective capabilities throughout language, photos, actions, and private context. Our fashions have been created with the aim of serving to customers do on a regular basis actions throughout their Apple merchandise, and developed responsibly at each stage and guided by Apple’s core values. We sit up for sharing extra data quickly on our broader household of generative fashions, together with language, diffusion, and coding fashions.

[1] We in contrast in opposition to the next mannequin variations: gpt-3.5-turbo-0125, gpt-4-0125-preview, Phi-3-mini-4k-instruct, Mistral-7B-Instruct-v0.2, Mixtral-8x22B-Instruct-v0.1, Gemma-1.1-2B, and Gemma-1.1-7B. The open-source and Apple fashions are evaluated in bfloat16 precision.

Leave a Reply

Your email address will not be published. Required fields are marked *