<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/e58510ef-3ffb-4a09-9bfb-4329cb176596.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/e58510ef-3ffb-4a09-9bfb-4329cb176596.jpeg"/>
Author | Fan Xinru
Editor | Zhou Zhiyu
The edge-side model has become a battleground for technology giants. Following Google, Samsung, and Microsoft, Apple has also joined the fray.
In the early hours of April 25th, Apple released OpenELM. This is a brand new open-source language model (LLM) series that can run entirely on local devices without the need to connect to cloud servers.
In the past, Apple's ecosystem was relatively closed. Now, with an open attitude, it has joined the wave of open-sourcing large models, hoping to catch up with the AI trend.
Perhaps at the Worldwide Developers Conference (WWDC 2024) in June, the outside world will be able to see how Apple innovates in AI technology through OpenELM and integrates it into devices like the iPhone to improve efficiency.
This will also be Apple's comeback battle after years of criticism from investors who believed it was &#34;lagging behind&#34; in AI and large models. With Apple's own chip resources and integrated hardware and software ecosystem, the seemingly &#34;lagging&#34; Apple may also have the opportunity to overtake.
<h2>Encirclement</h2>
During the investor conference call in February this year, Tim Cook explained Apple's &#34;lagging behind&#34; in the field of artificial intelligence.
&#34;Our model has always been to do the work first and then talk about it, rather than being ahead of the game.&#34; At the meeting, Tim Cook revealed that Apple is doing a lot of work in AI internally and will discuss it later this year.
However, before the &#34;later this year&#34; arrived, Apple released its own open-source language model, OpenELM.
OpenELM is a series of open-source language models, divided into instruction fine-tuning and pre-training models, with four sets of parameters: 0.27B, 0.45B, 1.08B, and 3.04B, providing functions such as text generation, code, translation, and summarization.
Compared to most models around 8B, its requirements for chips are also lower. In terms of large model capabilities, Apple stated that compared to the 1.2B parameter OLMo model, the 1.1B parameter OpenELM has an accuracy rate 2.36% higher, but only requires half the token amount for pre-training compared to OLMo.
Compared to Apple's previously announced MM1 multimodal large language model, the newly released OpenELM is not only smaller in size, but its most notable feature is that this series of language models can complete inference and fine-tuning on the edge without the need to connect to the cloud, specifically developed for mobile devices like phones.
The so-called edge-side large model typically refers to large models deployed and run on terminal devices (such as phones, tablets, smart speakers, etc.). Unlike models like ChatGPT, these large models can perform inference tasks and fine-tuning on terminal devices without being limited by the network
Actually, almost when large models became a popular term in the tech industry, discussions on how to implement large models at the edge, especially on smartphones closely related to people's daily lives, have been emerging one after another. However, constrained by the size of large models and factors such as chip performance and energy consumption, cars were once seen as the most likely terminal for running edge models.
However, after more than a year of exploration, with the continuous improvement of large model performance, large models with fewer parameters and stronger performance have begun to appear, providing the possibility for models to be implemented at the edge and making AI Phones carrying edge models the focus of pursuit for various tech companies.
Today, mainstream smartphone manufacturers have almost all launched their own large edge models. For example, OPPO's AndesGPT, vivo's BlueHeart large model, Xiaomi has the large model MiLM, and Samsung has also launched Galaxy AI.
Recently, major tech companies have intensively released a wave of large edge models.
Just 2 days before the release of OpenELM, on April 23rd, Microsoft open-sourced Phi-3-mini, a small language model that can be deployed at the edge on its official website. Earlier, Meta released its latest open-source model Llama3, providing pre-training and instruction fine-tuning versions of 8B and 70B. Subsequently, Qualcomm announced a collaboration with Meta to optimize the performance of Llama3 on edge devices such as smartphones and PCs. In China, SenseTime has just released the 1.8B SenseChat-Lite version of the large edge model.
This time, Apple seems to be late again.
<h2>Breakthrough</h2>
In recent years, Apple has rarely been able to launch disruptive products, with the iPhone becoming the biggest contributor to Apple's revenue in all product lines. Apple's financial data for the first quarter of 2024 showed that Apple's revenue in the fourth quarter of 2023 was $119.6 billion, a 2% year-on-year increase; profit was $40.3 billion, a 12% year-on-year increase. Among them, the iPhone contributed $69.7 billion, accounting for over 58%.
However, in recent years, iPhone sales have also faced challenges. On April 25th, the latest quarterly tracking report on smartphones released by the International Data Corporation (IDC) showed that in the first quarter of 2024, the overall shipment volume of smartphones in China was about 69.26 million units, a 6.5% year-on-year increase, performing better than expected. At the same time, iPhone sales in China decreased by 6.6% year-on-year, and its market share also dropped to 15.6%, tied with OPPO at 15.7% for third place. Honor and Huawei tied for the top two in shipment volume.
Wang Jiping, Vice President of IDC China, believes that Honor's rise to the top is driven by AI as a key growth engine. Thanks to AI functions, the new flagship Magic6 series from Honor had a first-quarter shipment volume exceeding the sum of the previous two quarters of the previous generation products.
As various smartphone manufacturers are entering the era of AI phones, Apple, under pressure, also urgently needs to join the war of large models to win market confidence
However, Apple still has a chance.
Nowadays, for some smartphone manufacturers, simply &#34;squeezing&#34; large models into phones seems to no longer be the biggest challenge. The problem lies in how to use large models to provide users with a differentiated, personalized, and even disruptive experience. This not only tests the capabilities of on-device large models, the performance of phone chips, and power consumption control, but also requires deep integration of large models with phone software and hardware systems, permeating all aspects of phone applications.
In this regard, choosing to develop in-house chips and emphasizing software-hardware integration from the beginning, Apple, with a complete development ecosystem, clearly has an advantage.
Currently, whether domestic or international major players, most of the exploration of AI phones is carried out through the collaboration of large model vendors, chip manufacturers, and phone manufacturers. For example, Meta chose to cooperate with Qualcomm to explore the possibility of deploying large models like Llama3 8B and 70B on the device side. Earlier, Alibaba Cloud's large model Tongyi Qianwen needed to collaborate with MediaTek to adapt to mobile devices. And most phone manufacturers do not have in-house chips and need to rely on chip suppliers. This means additional customization costs and adaptation and optimization costs.
At the same time, Apple, with its in-house chips, large models, and software-hardware ecosystem, can avoid many of these troubles.
For example, the benchmark test of OpenELM was conducted on Apple devices. Official papers show that the benchmark test of OpenELM uses a MacBook Pro equipped with the M2 Max chip, 64GB of memory, and running macOS 14.4.1. Furthermore, Apple has also released the code to convert models to the MLX library, allowing technical personnel to perform inference and fine-tuning of large models on Apple devices.
This means that OpenELM can smoothly integrate into Apple's entire ecosystem, bridging the gap between chips, phone hardware, and software layers, and better adapting and optimizing large models at the system level. This is an advantage that other phone manufacturers do not possess.
The software-hardware integrated strategy has kept the iPhone at the forefront of the mobile industry for many years. In recent years, Apple's persistence in in-house chips and software-hardware ecosystem integration may help Apple seize an opportunity to overtake in the era of large models. From the iPhone to the AI Phone, it seems that the distance is not far away

AAPL

Apple releases the end-side large model OpenELM

Apple's AI Phone Strikes Back