Xiaomi’s MiMo-V2.5-Pro UltraSpeed: 1,000+ TPS on 1T MoE Model via Standard 8-GPU Nodes

Xiaomi has unveiled MiMo-V2.5-Pro UltraSpeed, claiming a breakthrough inference speed of over 1,000 tokens per second (tps) for a 1-trillion parameter (1T) Mixture-of-Experts (MoE) model. Remarkably, this performance was achieved on a standard 8-GPU commodity server, rather than specialized wafer-scale or high-SRAM hardware like Cerebras or Groq.

• ▶ Software-Defined Performance: Xiaomi is challenging the dominance of specialized AI ASICs by proving that commodity GPUs, when paired with elite-tier software optimization, can deliver world-class throughput.
• ▶ The TCO Revolution: Achieving 1k+ TPS on standard hardware suggests a massive reduction in the Total Cost of Ownership for 1T-scale models, shifting the barrier to entry from custom silicon to software stack efficiency.

Bagua Insight

This is a “shots fired” moment for the inference market. By hitting these metrics on standard H100/A100 clusters, Xiaomi is effectively commoditizing high-speed, large-scale inference. The competitive moat is shifting from hardware availability to the depth of the software stack—specifically in kernel fusion, memory management, and MoE routing efficiency. If verified, this achievement threatens the premium positioning of AI hardware startups that rely on specialized architectures. Xiaomi is signaling that it is no longer just a consumer electronics giant but a hardcore AI infrastructure player capable of out-engineering the industry at the lowest levels of the stack.

Actionable Advice

Infrastructure leads should re-evaluate their hardware roadmaps; specialized AI chips may no longer be the only path to ultra-low latency for massive models. Engineering teams should prioritize MoE-specific optimizations and advanced quantization techniques to maximize existing GPU ROI. The focus must shift from “more GPUs” to “smarter kernels.”