Xiaomi's O3 flagships bring agent-class AI to phones and cars
Xiaomi unveiled three self-designed XRING chips at a Beijing technical briefing today, and while the flagship phone silicon made the news, the more telling move is the board stretching from handset SoCs into cars and on-device agents. The company is positioning its own silicon — not just its phones — as the compute backbone of a full pocket-to-car AI ecosystem.
The centerpiece is the 玄戒O3, Xiaomi's second flagship phone SoC and the industry's first to break five million on AnTuTu. The 24-billion-transistor, 3nm design pairs a 10-core all-big-core CPU running to 4.35 GHz with a 200-TOPS four-core NPU tuned for running large language models locally — supported by a new SIMT vector unit and quantization work shared with Xiaomi's own MiMo models, which Xiaomi says makes its on-device inference roughly 45 percent faster than mainstream flagship chips. It also carries a thoroughly reworked security stack with EAL5+ certification. The O3 ships this September inside the Xiaomi 18 Fold.
What makes this launch bigger than a phone chip is where the other two pieces go. The 6nm O100, which we covered this morning — Xiaomi's XRING O100 brings 330 tokens-per-second LLMs to devices — is a wafer-on-wafer AI accelerator reaching 1.22 TB/s of memory bandwidth for on-device agents. The D100, billed as China's first 3nm automotive AI chip, pairs a 20-core CPU with a 16-core NPU and supports up to 160 GB of memory — enough to run a 200-billion-parameter model inside a car fully offline, or to be linked across multiple chips for heavier local workloads. Both the O100 and D100 are developed and due to reach commercial products next year.
Why this matters: Xiaomi is betting that the winning layer of the agent era is local, not cloud. Running an agent-grade model on device sidesteps both per-token bills and the latency and privacy exposure of a round-trip to a data center — the same argument that has pushed Apple, Qualcomm and MediaTek into on-device silicon. Xiaomi is effectively trying to own the hardware layer of that shift across the whole consumer ecosystem, from pocket to steering wheel. As with any vendor launch event, these are the company's own numbers; the 200 TOPS, the five-million score and the 200-billion-parameter claim all deserve independent verification before being taken at face value.
What to watch: whether the O3's AnTuTu headline survives third-party reruns, and whether the D100's offline-200B promise translates into a real in-car product rather than a spec-sheet ambition.
Do on-device chips genuinely beat the cloud for everyday agent tasks, or is local AI mostly a privacy story? Tell us in the comments.
