Xiaomi Ring O3 is here. Xiaomi rolls out three core offerings all at once, ushering in the grand convergence of "Chips + OS + AI".

459 days ago, Xiaomi released Xuanjie O1, making its return to the arena of flagship mobile SoC. 459 days later, the Xuanjie O3 is not only officially launched, but also comes with more "siblings".

On the afternoon of August 24, Xiaomi held the Xuanjie Chip Technology Communication Conference. Zhu Dan, head of the Xuanjie chip product line, unveiled three self-developed chips at one go: the Xuanjie O3 for flagship smartphones, the Xuanjie O100 AI acceleration chip specially designed for edge-side large models, and the Xuanjie D100 high-computing intelligent driving chip for automotive scenarios.

The Xuanjie O3 will make its debut on Xiaomi's new-generation large foldable 18 Fold, which is officially released in September. The other two chips have finished R&D at present and are scheduled to be officially commercialized next year.

Image source: Xiaomi

According to Lu Weibing's previous disclosure, the cumulative shipment of Xuanjie O1 on three end products has exceeded 1 million units, and the Xuanjie product line will maintain an annual iteration speed. Lei Jun also stated that Xiaomi restarted its large-scale chip R&D plan in 2021, planning to invest at least 10 years and no less than 50 billion yuan. As of April 2025, the cumulative R&D investment in the Xuanjie series has exceeded 13.5 billion yuan.

The launch of O3, O100 and D100 undoubtedly proves that Xiaomi is still firmly investing in the path of self-developed chips. But on the other hand, a single O3 obviously cannot solve all related problems, which requires more flexible technical routes and more Xuanjie chips, especially under the background that "Xiaomi Xuanjie" is expected to become the AI computing power base of "Xiaomi's full ecosystem of people, vehicles and homes".

However, for both users and Xiaomi, Xuanjie O3 is still the most important chip today, and also the core focus of this article from Lei Tech (ID: leitech).

Upgrades of Xuanjie O3: Not just larger or faster

Judging from the announced specifications, Xuanjie O3 is still a quite aggressive flagship SoC, which continues to adopt 3nm process. The number of transistors has increased from 19 billion of O1 to 24 billion, and the chip area reaches about 133mm².

Xiaomi did not rush to reduce the chip area and cost, but still chose to leave enough space for performance improvement.

The change of CPU is more obvious. Xuanjie O1 adopts a 10-core 4-cluster design, including Cortex-X925 super core, A725 performance core and A525 efficiency core. For O3, Xiaomi redesigned the entire CPU combination, adopting Arm's new-generation C1 series IP, forming a 10-core full-performance-core architecture composed of 6 super cores and 4 large cores, with a maximum main frequency of 4.35GHz.

According to the data given by Xiaomi, the CPU performance of Xuanjie O3 is 60% higher than that of O1, and the power consumption under medium and low loads is reduced by 25%.

Image source: Xiaomi

In the past, the common idea of mobile SoC was to let small cores be responsible for background and light loads, and then use large cores and super cores to take over heavier tasks step by step. But from Arm C1 to recent generations of Dimensity chips, mobile processors are trying to cover a wider performance range with more energy-efficient large cores.

Xiaomi has obviously made its own choice. The code previously dug out from Mi Code has revealed in advance that Xuanjie O3 with code name "lhasa" will change from the 4-cluster design of O1 to a 3-cluster design, and the core frequency has also been fully increased. Looking back now, those codes have already outlined the basic features of O3 in advance.

The GPU of O3 has also been greatly upgraded. It is the first to be equipped with 16-core G2-Ultra NX graphics processor, which has 85% higher performance than Xuanjie O1, while the power consumption is reduced by 64%.

However, the peak figures at the time of release can only be used as reference. The AnTuTu V11 score announced by Xiaomi reaches 5228014 points, and the test environment is clearly marked as a low-temperature laboratory. For a mobile SoC, how fast it can run in actual use still depends on how much performance it can retain under normal temperature, normal power consumption limit and continuous load after being installed in the mobile phone.

These indicators still need to be tested after the Xiaomi 18 Fold and Xiaomi Pad 9 Pro Max are launched on the market.

Image source: Xiaomi

There is another upgrade on Xuanjie O3 that is easily covered by CPU scores - the memory.

As the industry's first SoC supporting LPDDR6, the Xuanjie O3 has a rate of 10667Mbps, and adopts a 4×24bit, total 96bit memory bit width design, with a bandwidth of 113.8GB/s, which is 48% higher than that of Xuanjie O1.

For today's flagship SoC, CPU needs data, GPU needs data, image ISP needs data, and large models are famously heavy on memory bandwidth. If more computing units are added to the chip but the memory and on-chip interconnection cannot keep up, the situation that computing power is waiting for data will occur.

Therefore, O3 has also redesigned a set of "unified fusion bus architecture". According to Xiaomi, this protocol can open up the data link inside the chip, and the overhead caused by protocol conversion can be reduced by up to 75%.

Image source: Xiaomi

Computing IP such as CPU and GPU can come from Arm, but to connect dozens of different modules and allow CPU, GPU, NPU, ISP and memory controller to efficiently exchange data within limited power consumption and area, chip manufacturers need to handle NoC, Cache, QoS, scheduling and bus design on their own.

As Xuanjie has developed to the second generation, Xiaomi has begun to put more energy on this part, and this upgrade is also a microcosm of the transcendence over conventional performance upgrades from O1 to O3.

O3 has a new title: AI flagship SoC

At this technical communication conference, Xiaomi's official positioning of Xuanjie O3 is "AI flagship SoC", and the listed chip evolution direction is very straightforward:

High energy efficiency, high bandwidth and high computing power.

Image source: Xiaomi

To trace the root cause, Xuanjie O3 is ultimately aimed at "building an AI computing power base for Xiaomi's full ecosystem terminals of people, vehicles and homes".

When O1 was released, what the outside world discussed most was the 3nm process, CPU, GPU, and whether Xiaomi could successfully make a flagship mobile processor again. For O3, Xiaomi has begun to integrate Xuanjie into a larger technical blueprint.

Xiaomi has actually talked about this blueprint for more than a year.

As early as March 2025, Lu Weibing clearly stated that AI, OS and chips are the three long-term core technologies of Xiaomi. At that time, Xiaomi planned to invest about a quarter of its R&D funds, 7 billion to 8 billion yuan in AI, covering AI infrastructure, language models, multimodal models, super Xiao Ai, smart cockpit and intelligent driving and other directions.

In January this year, Lei Jun gave a more specific goal:

In 2026, Xiaomi expects to realize the "great integration" of self-developed chips, self-developed OS and self-developed large AI model on one of its end products.

Today, all three lines have made great progress respectively. For chips, there is the newly announced O3; for operating system, it has developed to HyperOS 4; for large models, there is the MiMo series.

At Xiaomi's Q2 earnings call a few days ago, Lu Weibing explained this matter more clearly. He divided Xiaomi's current AI layout into several lines, the bottom layer is the basic large model MiMo, which extends to mobile phones, homes, automobiles and robots.

On the mobile phone line, he clearly proposed to comprehensively reconstruct the traditional Android with AI, and finally evolve to "AIOS". HyperOS 4 is already one of the steps, which has begun to obtain execution capabilities through intelligent agents.

Xiaomi CFO Lin Shiwei revealed that the weekly and monthly global Token call volume of MiMo-V2.5 once ranked first in July this year. Xiaomi is still training new models and plans to bring MiMo to PCs. At this stage, Xiaomi is not in a hurry to require its AI business to make profits. The direction given by the management is still to continue investing, so that AI can reconstruct the "full ecosystem of people, vehicles and homes".

Image source: Xiaomi

Looking back at Xuanjie O3, we can understand why Xiaomi now emphasizes "high bandwidth" and "AI computing power base". After large models are deployed to end devices, the requirements for mobile chips have gone far beyond the number of TOPS of NPU.

The operation of models involves memory bandwidth, cache, operator support, quantization accuracy, and power consumption scheduling; the long-term operation of Agents will involve system permissions, background mechanisms, and calls between different applications. If the chips, systems and models come from different companies, each layer needs to adapt within the boundaries set by others.

When Xuanjie, HyperOS and MiMo are all mastered by Xiaomi itself, Xiaomi has the opportunity to carry out real independent design in reverse:

• For any operator that MiMo needs, the NPU can be optimized in a targeted manner;
• If edge-side models require higher bandwidth, the memory system can be adjusted;
• If Agents need to run for a long time, modifications can be made from the low-power scheduling of the chip all the way to the OS background mechanism.

This is also a path that Apple has proven effective repeatedly over the past decade: chips, operating systems, software and hardware are integrated into the same R&D system, so that product teams can solve problems from the bottom level.

Xiaomi's situation is of course somewhat different. In addition to mobile phones, tablets, watches, home appliances, automobiles, and the robots that are under development, a single Xuanjie O3 obviously cannot solve all problems, so O100 and D100 were launched.

Image source: Xiaomi

Xuanjie O100 is Xiaomi's first high-bandwidth AI acceleration chip specially designed for edge-side large models. Through 6nm wafer-level vertical stacking and Hybrid Bonding technology, the computing and storage units are placed closer to each other, finally achieving a memory bandwidth of 1.22TB/s, which is about 16 times that of traditional flagship mobile phones, and the maximum reasoning speed of edge-side large models reaches 330TPS.

Xuanjie D100 is the first 3nm intelligent driving chip in China, equipped with 20-core high-performance CPU and 16-core high-computing NPU, supporting a maximum of 160GB memory, and can locally deploy large models with a maximum parameter scale of 200B.

The three chips handle three different types of workloads, but all point to the same goal: Xiaomi hopes to take more and more AI computing power of the "people, vehicles and homes" ecosystem into its own hands.

Xiaomi has even made an AI Cube engineering machine, which is equipped with O3, O100 and D100 chips at the same time, for local deployment of 120B/3B large models. It has no launch plan for the time being, and is more like a "demonstration field" for Xiaomi to verify the cooperative operation of the three computing platforms.

Final notes

When Xuanjie O1 was released, the focus of everyone was: Xiaomi has successfully made a flagship SoC again.

A year later, Xiaomi 15S Pro and two Xiaomi Pad 7 series products have helped Xuanjie O1 achieve shipments of more than 1 million units. For Apple, Qualcomm and MediaTek, this is just a small number. But for Xiaomi, which has just re-entered the flagship SoC track, it is of extraordinary significance.

However, when the new generation of self-developed Xuanjie chips such as O3, O100 and D100 are about to officially integrate with HyperOS and MiMo large models, Xiaomi is facing a whole new set of examination questions - the ultimate test around the "full ecosystem of people, vehicles and homes".

We won't get the answer today. At least after Xiaomi widely applies Xuanjie to its full ecosystem terminals, can we evaluate the "real performance" of Xiaomi's long-term technology strategy to a certain extent. But Xiaomi has already brought a lot of surprises, and I don't mind having more expectations for them, just like the brand slogan Lei Jun set for Xiaomi:

Always believe that something wonderful is