Source | Bohu Finance (bohuFN)

Two years ago at WWDC, Apple made a big splash.

Apple proudly unveiled 'Apple Intelligence,' showcasing to the world what an AI smartphone looks like:

Siri will possess capabilities such as personal context awareness (reading emails, texts, calendars, and other personal information to answer questions), screen content awareness (understanding the content currently displayed on the screen and performing operations), and cross-application operations.

However, the new Siri not only failed to launch with iOS 18 but was repeatedly delayed by Apple until 2025 due to 'internal technical issues.' Craig Federighi, Senior Vice President of Software Engineering, later admitted to The Wall Street Journal that during internal testing, the prototype 'continued to produce an excessive number of unreliable results,' with the system's success rate ranging from only 60% to 80%, failing to meet Apple's product standards.

Due to the delayed implementation of AI features, Apple faced multiple class-action lawsuits and had to agree to pay $250 million to settle with iPhone purchasers.

This gap was only filled this year.

A few days ago at WWDC, Apple teamed up with Google to reconstruct the Apple Intelligence architecture based on the Gemini family. On top of the new Apple Intelligence architecture, Siri AI with contextual understanding and screen awareness, along with multiple system-level AI features, were introduced.

According to Apple, this represents 'a qualitative leap forward.'

However, these updates did not receive positive feedback, and Apple's stock price fell after the event, with its market value evaporating by over $230 billion. Ultimately, there was no surprise.

Over the past fifteen years, Tim Cook has built upon the solid foundation of 'integrated hardware and software' created by Steve Jobs, increasing Apple's market value from $350 billion to $4 trillion. However, the industry is now, to some extent, on the eve of a transition from feature phones to smartphones—as the most important mobile terminal, more and more AI companies are attempting to reshape smartphones with AI. Doubao has already set a precedent, and OpenAI is planning to follow suit. With its home almost being infiltrated, Apple has not delivered sufficiently impressive innovations in AI.

So, the question arises: Is Apple's first answer sheet for AI smartphones really that bad?

01 The Only Black Mark is Google

If one must find a black mark in this update, it is probably that Apple finally admitted it couldn't handle model capabilities on its own and chose to collaborate with Google. The new generation of Apple Foundation Models is based on the Gemini family's collaborative models, with two on the device side and three in the cloud.

On the device side, there is the AFM 3 Core with 3 billion parameters for daily lightweight tasks and the more powerful AFM 3 Core Advanced with 20 billion parameters, a sparse model. In the cloud, there are the AFM 3 Cloud, the main server model; the ADM 3 Cloud, specialized in image generation and editing; and the most powerful AFM 3 Cloud Pro.

Although Apple executives said:

'We use zero Google Assistants.'

'All these models are custom-built for Apple Silicon, trained with proprietary data, and fine-tuned using the output of Gemini's cutting-edge models.'

You can basically understand that the foundation of Apple's new models is Gemini.

Apart from that, Apple has done well in other areas.

In the cloud, Apple has consistently adhered to its privacy policy. Apple has built a dedicated cloud computing infrastructure for AI—Private Cloud Compute—where data remains within the domain and is end-to-end encrypted.

Regarding the underlying models, according to previous leaks, Apple's agreement with Google is that Apple pays $1 billion annually to obtain a customized version of the 1.2 trillion-parameter Gemini model. Users can use this customized model for free. Apple has also public (made public) third-party adaptations, including many familiar names like Claude, ChatGPT, and the native Gemini. Users need to pay for these themselves.

In terms of AI capabilities, although Apple's AI is not outstanding, it is on par with the experiences offered by other competitors.

According to the artificial intelligence terminal standards recently introduced in China, terminal intelligence is divided into four levels: L1 Response Level, L2 Tool Level, L3 Assistance Level, and L4 Collaboration Level. L1 can only understand simple instructions, L2 can complete simple multi-step operations, L3 can understand complex intentions and proactively serve users, and L4 Collaboration Level will be refined in subsequent revisions based on industry development.

If we consider proactive service as the dividing line between L2 and L3, most manufacturers, including Apple, are at the L2 level.

The Doubao smartphone was previously seen as an innovative example of AI smartphones. Its GUI operation capability allows it to simulate user clicks, and even complex tasks requiring dozens of steps can achieve a relatively high success rate.

Apple's Siri AI is also evolving towards an Agent direction, possessing capabilities such as screen awareness, personal context understanding, and task execution.

On the one hand, as a highly integrated system-level AI, Siri AI naturally possesses cross-APP execution capabilities similar to Xiaomi's Claw. For example, when you browse photos, you can directly ask Siri AI where this place is, and it will identify the location for you. When you want to ask how to get there directly, Siri AI can also call up the Messages and Maps apps to generate navigation for you.

On the other hand, Siri AI possesses personal context capabilities. Not only can users use and view historical conversation records on multiple devices with the Siri AI as a standalone app, but it can also extract information from your emails, texts, photos, etc., to respond to your needs.

In addition, system applications such as Safari, Messages, Mail, Calendar, Phone, and Photos have all added AI capabilities. The most typical example is that the Photos app can use AI and 3D modeling technology to generate new perspectives from existing photos.

Nicole Peng, Vice President of Consumer Business Research at Omdia (formerly Canalys), also stated that in terms of actual AI capabilities, Apple is at least in the first tier when compared horizontally with the AI capabilities bring your own ( bring your own , meaning ' bring your own ' or 'inherent') by other smartphone manufacturers in the market.

02 Advantages of Integrated Hardware and Software in the AI Era

It is reported that after launching the Doubao smartphone last year, ByteDance accelerated the advance ( advance , meaning 'promotion' or 'advancement') of its AI smartphone project in 2026. The Ocean team responsible for the Doubao smartphone is ByteDance's core hardware team, Flow, which is on par with Douyin (TikTok) within ByteDance.

In addition to the Doubao smartphone, the Ocean team has also integrated ByteDance's hardware resources accumulated over the years, including but not limited to the original Smartisan smartphone, VR headset PICO, and smart earphones Oladance.

ByteDance's emphasis on terminals highlights the importance of intelligent terminals in the AI era. The ultimate destination for all consumer electronics products is integrated hardware and software. Any so-called form of empowerment or cooperation can only yield temporary advantages.

This is the biggest advantage for smartphone manufacturers in the AI era. Both Xiaomi and Apple are clear beneficiaries.

For example, Xiaomi can now connect smartphones and IoT devices. Miclaw can read and understand the information on your smartphone, provide services similar to a 'butler,' and also coordinate tasks across the Xiaomi ecosystem: when you say to Miclaw, 'I'll bring friends home in half an hour; prepare the house,' it will break down and execute tasks: automatically coordinating Mi Home devices to adjust the lighting to warm tones, close the curtains, set the air conditioner to 26°C, start the air purifier, and play background music through the speaker.

A unified Language Model protocol and a complete scheduling framework from the device side to the cloud allow Apple to connect 2.5 billion devices. On Macs and iPads, you can evoke ( evoke , meaning 'invoke' or 'activate') Siri directly from Spotlight or select 'Ask Siri' from the right-click menu. In Vision Pro, Siri can even understand your gaze focus, knowing which object you are looking at, and finally provide information.

However, overall, Apple's unique integrated hardware and software capabilities have allowed it to accumulate more advantages in the AI era.

From a model perspective, the most powerful device-side model, AFM 3 Core Advanced, has up to 20 billion parameters. Logically, it would be difficult to fit a model of this scale into a smartphone, but Apple has achieved this through Instruction-Following Pruning. It places the complete model in flash memory (NAND) and keeps a small group of 'always-activated shared experts' in DRAM, only bringing in the corresponding experts when selected by the predictor. This way, the actual parameters mobilized each time are only between 1 billion and 4 billion.

The iPhone is also equipped with the first dynamically sparse LLM for large-scale consumer mass production ( mass production , meaning 'mass production').

In addition, Apple has also started to shift its chip design. According to a teardown analysis by Counterpoint Research, the Apple M5 Pro chip adopts a chipset-like structure, enhances the role of the GPU in artificial intelligence computing, and achieves the highest memory bandwidth to date among Apple's professional-grade system chips.

Memory bandwidth is crucial for on-device artificial intelligence. Because running large language models and professional AI workflows on devices requires not only computing power but also sufficient memory capacity and bandwidth to efficiently feed data to the models.

This indicates that Apple is adjusting its Mac chip strategy to actively adapt to the demands of the AI era.

Reference Sources:

1. EDA365 Electronics Forum: Significant Changes in Apple's Chip Strategy

2. Letter AI: A Transformed New Siri Turns the iPhone into a Doubao Smartphone

3. Jiemian News: A Comprehensive Look at WWDC2026: Cook's 'Swan Song,' Major AI Updates from Apple

4. TMTPost AGI: Apple's AI 'Turnaround' Still Relies on Gemini as the 'Brain' | WWDC26

5. APPSO: The iPhone Transforms into an AI Phone Overnight, but the Future of AI Smartphones Lies Beyond the Phone

The cover image and accompanying illustrations in this article are owned by their respective copyright holders. If the copyright holders believe that their works are not suitable for public browsing or should not be used free of charge, please contact us promptly, and our platform will make immediate corrections.