Xiaomi has unveiled the XRING O100, an AI accelerator designed for large models, at a chip technology briefing on August 24.

The company said the chip uses a 6nm 3D wafer-level stacking design and delivers bandwidth of 1.22TB/s. It adopts a near-memory AI computing architecture and Xiaomi's high-bandwidth matrix bus to support multicore parallel computing.

Running Xiaomi's MiMo 3B model, the XRING O100 can reach an inference speed of 330 tokens per second, according to the company.

The launch adds to Xiaomi's push to develop more of its AI and device computing stack in-house, with the XRING O100 expected to be deployed next year, according to the company.