How Xiaomi's AI phone reached 303 tokens/s on-device
The short version: Xiaomi demonstrated MiMo 3B running fully offline at about 303 tokens/s, with a reported 330 tokens/s peak and roughly 0.45-second time to first token. That speed came from a dual-chip, actively cooled engineering prototype—not a normal retail phone.
These are Xiaomi briefing and press hands-on figures, not an AICanRun benchmark. Model prompt, context, quant file, sustained duration, and per-run logs were not published, so the result is not independently reproducible yet.
What was actually inside the demo phone?
The prototype paired the XRING O3 phone SoC with a separate XRING O100 high-bandwidth AI accelerator. O100 vertically bonds two AI-memory wafers to an NPU wafer, shortening the data path and raising bandwidth to 1.22 TB/s. The chassis used an active fan with up to 10W of cooling; its cameras were removed and the internal layout was rebuilt around the accelerator.
This is not the retail Xiaomi 18 Fold
Xiaomi has announced that the September-bound Xiaomi 18 Fold will use XRING O3. It has not said the retail phone includes O100, active cooling, the prototype's memory arrangement, or its 303 tokens/s result. Physical RAM tiers are also still unannounced. We will not turn the retail phone into a verified device in our compatibility database until Xiaomi publishes those specifications.
Why memory bandwidth changed the experience
During autoregressive decoding, model weights must be read repeatedly for every generated token. That makes bandwidth—not only headline TOPS—a hard limit. O100's 1.22TB/s on-package memory and the compressed MiMo weights attack the same bottleneck from both sides. The result is far beyond the decode rate of today's phone LPDDR interfaces, but it also depends on specialized hardware and a model build consumers cannot download today.
O3 by itself remains interesting: Xiaomi reports 40% faster first-token processing, 45% faster decoding, and 26% lower power for its optimized MiMo 3B path. Those are relative figures; Xiaomi has not published an O3-only absolute tokens/s result that can be compared directly with the dual-chip demo.
The desktop sibling: Xiaomi AI Cube
Xiaomi separately showed an Android desktop prototype with 80GB of unified memory, a 150W power target, and local 3B + 120B fast/slow model routing. A coding demo generated and ran a Web Audio piano app locally. Xiaomi did not disclose the 120B model name, quantization, context, or generation speed, so that demo proves the product direction—not a reproducible benchmark.
For comparison, the open GLM-4.5V 106B model has a real 63.6GB Q4 GGUF and needs a 96GB-class system for a practical local setup. That gives useful scale to the Cube's 80GB demo while avoiding the assumption that Xiaomi used GLM-4.5V.
Sources and evidence
FAQ
Did Xiaomi really run MiMo 3B at 303 tokens/s on a phone?
Yes. At Xiaomi’s technical briefing, a modified foldable engineering prototype generated about 303 tokens/s fully offline and peaked near 330 tokens/s. It used both an XRING O3 application processor and a dedicated XRING O100 AI accelerator, plus active cooling.
Is the prototype the retail Xiaomi 18 Fold?
No. The demo chassis was heavily modified: its cameras were removed, the motherboard was redesigned, an active fan was added, and the device could not fold normally. Xiaomi has confirmed O3 for the future Xiaomi 18 Fold, but not O100, its RAM tier, or the demo performance.
What model did Xiaomi run locally?
The phone prototype ran Xiaomi MiMo 3B with a custom five-value quantization and hardware Huffman compression. Xiaomi says the compression reduces memory-bandwidth demand by about 30%.
Can I buy this Xiaomi AI phone?
No. It is an engineering demonstration, not a retail product. The separate Xiaomi 18 Fold is scheduled for September, but final RAM, memory configuration, and on-device model performance have not been published.