Advertisements
In the ever-evolving landscape of artificial intelligence, the recent partnership between Apple and Alibaba proves to be both unexpected and entirely logicalAs speculation swirled around which tech giant would team up with Apple for its Chinese AI initiative, the answer emerged quietly from the shadows: Alibaba.
Reports from media outlets such as The Information indicate that Apple has selected Alibaba as its collaborator for launching AI functionality in the Chinese market.
This decision highlights a crucial moment in the selection process, where other contenders, such as DeepSeek, fell short despite their initial visibility and investment in innovative AI technologies.
Examining the capabilities of Alibaba's Qwen model reveals a striking level of technological prowess amidst various major players in the AI domain, particularly in China, where Alibaba often remains under the radar.
While it may seem counterintuitive, being understated can prove a significant advantage in an industry as frantic and rapidly changing as AIHere, a less flamboyant approach often translates into a sturdy, dependable foundation.
The strength of a company is often measured through its performance within the open-source community, which serves as a litmus test for technical competenceAlibaba ranks among the few firms capable of standing firm in this arenaThis assertion holds true as illustrated by the recent leaderboard of the leading open-source AI models provided by Hugging Face, where all ten top models are essentially derivatives built upon Alibaba’s Tongyi Qwen framework.
This leaderboard has quickly become one of the most authoritative sources for evaluating open-source models globallyIts metrics encompass crucial areas such as reading comprehension, logical reasoning, mathematical computation, and factual question answering.
Moreover, the significance of establishing a reputable technical brand continues to rise, with Alibaba's Tongyi Qwen model outshining peers like Llama to become the benchmark within the open-source community.
Alibaba’s Qwen model, which sets itself apart as the industry frontrunner in offering multi-modal capabilities across varied scenarios, accommodates developers and businesses with sizes ranging from a modest 1.5 billion parameters to an impressive 110 billion, thereby catering to diverse application demands.
In particular, the two models focused on visual understanding, Qwen-VL and Qwen2-VL, have amassed a global download count exceeding 32 million
Advertisements
Just a week ago, the Qwen2.5-VL received a substantial upgrade, reigniting enthusiasm within the open-source community.
When we let the data speak for itself, the Hugging Face 2024 leaderboard reveals that Qwen2.5-1.5B-Instruct garners a 26.6% share in downloads, marking it as the most favored open-source model among global developers.
Currently, the number of Qwen derivative models has surpassed 90,000, evolving into the most extensive collection of open-source models worldwide.
In the waves of AI inference development, Qwen has emerged as a powerful attractorRecently, a renowned AI figure, Fei-Fei Li, alongside her team, developed a model known as s1, which was trained using fewer resources and data, yet matched the performance of OpenAI’s o1-Preview in several capabilities thanks to Qwen-32B distillation.
Such accomplishments stem from the robust foundational strength of the Qwen model, which simultaneously endows Alibaba with the confidence to meet Apple's rigorous standards for AI functionalities.
Additionally, as a stalwart of China’s AI industry, Alibaba holds an undeniable edge over emerging startups; its deep understanding of local user needs and its acquaintance with data compliance policies equip it to facilitate rapid adaptation and efficient localization of Apple’s AI capabilities.
However, the outcome of this collaboration was not preordainedOver the past few months, Apple meticulously examined multiple contenders, including Tencent, ByteDance, and DeepSeek.
During a recent earnings call, Apple CEO Tim Cook praised DeepSeek for its innovative efforts in reducing the training and operational costs of its V3 modelHe emphasized that efficiency-driven innovations are commendable, a sentiment that reflects on DeepSeek’s contributions.
Nevertheless, the company’s exit from this partnership doesn’t come as a surprise
Advertisements
DeepSeek's limitations regarding team size and experience with large-scale clientele became apparent, particularly when users faced error messages like "server busy, please try again later." In contrast, Alibaba's technological infrastructure and capabilities offer Apple a significant advantage in localized computing support.
Furthermore, through this collaboration with Alibaba, Apple aims to unveil AI functionalities that resonate with Chinese users' preferences—such as enhanced voice recognition and visual understanding—thereby boosting its product competitiveness and addressing market challenges.
Apple's objective is clear: to partner with an entity that can harness user data and comprehend the underlying requirements of Chinese customers, making it an integral part of developing smarter Apple devices.
As Apple evaluated various AI models developed by Chinese firms, it expressed dissatisfaction with these models’ capabilities in interpreting user intent and integrating real user experiences into the generated responses.
Ultimately, the choice of Alibaba as a partner seems to answer a critical question: can the selected AI solution withstand meticulous scrutiny and operate effectively in demanding scenarios? This partnership stands as a strong endorsement of Qwen's capabilities.
The AI experience for iPhone users will undergo transformation post Qwen integration.
With the dawn of a transformative era in mobile technology, OPPO's Chief Product Officer, Liu Zuohua, has pointed out that AI phones could represent the third significant phase in mobile history, following basic phones and smartphones.
The evolution of AI phones can be classified into three stages: the application-level AI which enhances individual applications; the system-level AI that assimilates intelligence at the operating system layer; and, lastly, achieving a state where "AI is the system," making AI central to mobile interactions.
The emergence of AI applications on mobile devices has grown increasingly sophisticated, especially with the recent success of DeepSeek, prompting manufacturers to widely adopt a “cloud-edge collaboration” model
Advertisements
Under this framework, lightweight tasks are processed on-device while heavier computational tasks are offloaded to the cloud.
At Apple’s WWDC last year, similar hierarchical architecture principles were showcased, which included:
Local Models: These primarily comprise fine-tuned 3B small models dedicated to tasks like summarization that, bolstered by adapters, perform robustly.
Private Cloud Computing: Tasks that exceed the local model’s capabilities transition to the cloudApple prioritizes end-to-end encryption to safeguard user data and ensuring privacy.
Third-Party LLM: Models for general knowledge Q&A and conversational tasks, such as Siri, may access external models like ChatGPT.
In the collaboration with Alibaba, the Qwen model might be integrated into the iPhone ecosystem through two primary avenuesOne would be a plugin-like approach that provides comprehensive AI services akin to a third-party LLM.
This strategy aligns with Apple's integration of ChatGPT in international marketsHowever, its partnership with OpenAI is not exclusive, leaving potential for collaboration with other models like Google’s Gemini or Claude later on in the overseas iPhone setupConversely, more localized Chinese AI frameworks might also be introduced into the Chinese market.
The second avenue entails collaborating with Alibaba to deploy distilled versions of smaller modelsUnlike DeepSeek’s focus on extensive 671 billion parameter models, Qwen pursues a more inclusive strategyEven DeepSeek itself incorporates four out of its six distilled open-source models based on Qwen, showcasing the technological prowess of Alibaba.
Moreover, while DeepSeek has touted its Janus-Pro-7B multimodal model as surpassing OpenAI DALL·E 3, it currently does not support multimodal capabilities on the consumer-facing side
In contrast, Qwen stands strong, as the latest Qwen 2.5 Max ranks among the most competitive multimodal models available.
Noteworthily, significant strides have been made in model compression techniques in 2024. Chinese manufacturers have made remarkable progress in optimizing AI models, achieving substantial parameter reductions while ensuring output quality through advanced quantization and pruning techniques.
According to Canalys, for instance, Xiaomi’s MiLM2 saw its model parameters trimmed from 6 billion to 4 billion, while Honor and Vivo reduced theirs from 7 billion to 3 billionSimilarly, Tongyi Qwen managed to compress a 7 billion parameter model down to 2.2 billion while preserving commendable comprehension and generation abilities.
On Apple's side, tools such as Core ML also support quantization of model weights, reducing the default float 32 precision down to even lower bit depths, thus saving storage while improving inference stability and reliability.
The quantization process ensures that intermediate tensors typically maintain float 32 or float 16 precision and that weights are de-quantized to match the intermediate tensor precision during runtime, thereby improving inference accuracy.
Notably, last year, Apple achieved a groundbreaking experience on the iPhone 15 Pro through various optimization measuresThis included latency analysis and power usage checks, which led to a model yielding a response for the first token in just 0.6 milliseconds while producing 30 tokens per second for an exceptional user experience.
Ultimately, Apple’s goal remains to design a genuine system-level AI assistant.
While models serve as a critical entry point, it is the ecosystem that becomes the winning cardWith Alibaba's ecosystem encompassing e-commerce, payments, local services, and entertainment, a partnership with iPhone AI could enable users to seamlessly complete service bookings through simple voice commands and tap into enterprise resources from platforms such as DingTalk to further broaden application scenarios.
The rivalry among AI smartphones is transitioning from mere technological competition to a contest of ecosystem integration and user experience
Advertisements
Advertisements
Leave a Comment