Xiaomi’s New AI Chips Show What AI-Ready Silicon Looks Like

Xiaomi’s August 24, 2026 Xuanjie communication briefing was less a single-product launch than a roadmap for an AI computing stack. The company introduced the Xuanjie O3 for flagship phones and tablets, the O100 as a dedicated edge-AI accelerator, and the D100 for intelligent-driving workloads. It also showed an AI Cube prototype designed to run large language models locally using multiple Xuanjie chips. 156

The announcement’s most important idea was architectural: improving local AI requires more than adding theoretical compute. Xiaomi is targeting the movement of model weights and intermediate data—the bottleneck often called the memory wall—while leaving a separate software and product challenge unresolved: the application wall.

Xiaomi’s three chips at a glance

Chip	Role	Key reported specifications	Manufacturing and timing
Xuanjie O3	Flagship smartphone and tablet SoC	3nm process, 24 billion transistors, claimed 200 TOPS of tensor performance and an AnTuTu score of 5.22 million	Reported to use TSMC production arrangements; mass production had begun and the chip is scheduled to debut in the Xiaomi 18 Fold in September 2026 248
Xuanjie O100	Dedicated edge-AI accelerator for local large-model inference	6nm process and 1.22 TB/s memory bandwidth	Uses a near-memory architecture with vertically stacked compute and memory; commercial deployment is planned for 2027 2569
Xuanjie D100	Automotive and intelligent-driving AI processor	3nm high-compute chip for autonomous-driving workloads	Reported to be part of Xiaomi’s TSMC manufacturing arrangements; commercial deployment is expected in 2027 245

The figures are primarily company-reported or reported by technology and news outlets. In particular, the O3’s 5.22-million AnTuTu result should be treated as a vendor-reported benchmark until independent testing is available.

Xuanjie O3: a flagship phone SoC built around local AI

The O3 remains Xiaomi’s conventional mobile platform in product terms: it is intended to power premium smartphones and tablets rather than operate as a standalone AI card. Xiaomi reported a 3nm design with 24 billion transistors, a 10-core CPU, a 16-core G2-Ultra NX GPU and an AnTuTu score of approximately 5.22 million. 41012

Its rebuilt NPU is the more consequential part of the announcement for on-device AI. Xiaomi rates it at up to 200 TOPS of tensor compute and says it is optimized for the company’s MiMo on-device models. 1818

O3’s compression and latency strategy

Xiaomi’s O3 design addresses AI efficiency through both data movement and compression. Reporting on the chip describes a reconstructed unified-fusion bus and physical routing that reduce memory-access latency to 82 nanoseconds in static tests and 177 nanoseconds with background load, compared with 121 and 384 nanoseconds respectively for the previous O1. 17

The chip also reportedly includes hardware Huffman lossless compression for Xiaomi’s five-value-quantized MiMo model. Xiaomi says that approach can reduce memory-bandwidth usage by 30 percent, alongside 28MB of NPU near-memory. 17

The practical point is that an AI processor can have abundant arithmetic capability and still feel slow if it cannot deliver model data quickly enough. Lower latency, near-memory storage and model-aware compression help reduce that friction without treating TOPS as the only measure of performance.

Xuanjie O100: a specialist accelerator for large models

The O100 is not a complete phone processor. It is a dedicated 6nm neural-processing and AI-acceleration chip intended to run large models at the edge, including Xiaomi’s MiMo models on consumer devices. 289

Its headline specification is 1.22 TB/s of memory bandwidth. That number matters because local large-model inference is frequently limited by how quickly the accelerator can access model parameters and intermediate results, not simply by how many operations its compute units can theoretically perform. 15

Near-memory computing attacks the memory wall

Xiaomi’s answer is near-memory computing. The O100 vertically stacks compute and memory using 3D wafer-on-wafer and hybrid-bonding approaches, shortening the physical path between the NPU and memory. 59

This is a different strategy from merely increasing TOPS. More TOPS can raise peak arithmetic throughput, but it does not automatically solve the bandwidth and latency costs of repeatedly moving large models. By placing compute closer to memory, the O100 is designed around the data-access problem itself.

Xiaomi positions the accelerator for local AI across phones, PCs, cars and robots. The company also used it in an AI Cube local-LLM prototype, alongside other Xuanjie chips, to demonstrate a multi-chip approach to edge inference. 1515

The O100 had completed research and development but was not the chip slated for the Xiaomi 18 Fold. Reports put its commercialisation timeline in 2027. 26

Xuanjie D100 extends the strategy into cars

The D100 is Xiaomi’s 3nm high-compute AI processor for intelligent driving and autonomous-driving workloads. Its role is distinct from both the mobile O3 and the general edge-inference focus of the O100: it is intended to supply AI compute for vehicles and help extend Xiaomi’s silicon strategy across its broader device ecosystem. 245

Some reporting describes the D100 as a 20-core processor supporting up to 160GB of unified memory and local processing of models with as many as 200 billion parameters. Those are reported specifications rather than independently verified performance results. 11

Like the O100, the D100 had completed development but was expected to reach commercial deployment in 2027. 256

Why these chips are AI-ready—but do not define an AI phone

On a narrow hardware benchmark, the O3 easily clears the often-cited 30-TOPS threshold associated with an AI phone. Its claimed 200 TOPS is more than six times that figure. 18

But a TOPS rating describes potential compute capacity, not what a person can actually accomplish. Xiaomi’s new silicon addresses several foundations of local AI:

• Compute: the O3’s NPU is rated at 200 TOPS.
• Bandwidth: the O100 is specified at 1.22 TB/s.
• Latency: the O3’s reported memory-access figures are substantially lower than those of the O1. 17
• Model efficiency: hardware compression and quantization reduce the amount of data that must move through memory. 17
• Local operation: the chips are designed to support inference on devices rather than relying exclusively on remote servers. 15

That makes the chips important infrastructure for AI devices. It does not establish a universally accepted definition of an AI phone. Reporting from the briefing indicates that Xiaomi did not reduce the category to a single chip specification and noted that product claims must comply with applicable requirements. The supplied reporting does not establish a more precise compliance rule, so the safest conclusion is that “AI phone” remains a combination of hardware capability, software integration, user experience and regulatory positioning—not a settled benchmark label. 14

From the memory wall to the application wall

The memory-wall problem is technical: large models require fast, efficient movement of data. O3’s latency and compression work, along with O100’s near-memory design, directly target that problem.

The application-wall problem is broader. A genuinely useful AI phone would need a trusted system-level agent that can understand context, retain appropriate memory, respect permissions and complete multi-step actions across services. A user might ask it to find information, update a calendar, prepare a document or coordinate a route; the value would come from completing the workflow across apps, not merely generating a response in a chat window.

That is the significance of the human–computer-interaction shift Xiaomi executives have pointed toward: AI must change how people use a device, rather than simply add a chatbot or produce local text more quickly. The O3 and O100 make more ambitious on-device agents technically more plausible, but the chips do not deliver cross-app autonomy by themselves.

Xiaomi’s August briefing therefore showed a credible foundation for AI-capable phones, edge terminals and intelligent vehicles. The next test is not whether the silicon can process a larger model. It is whether Xiaomi can turn that capability into a reliable, permission-aware system that moves from user intent to completed action.