MiMo-V2.5-Pro-UltraSpeed: Xiaomi Hits 1,000 Tokens Per Second on a Trillion-Parameter MoE — On Commodity GPUs

Xiaomi and TileRT released MiMo-V2.5-Pro-UltraSpeed on June 8, 2026 — a 1-trillion-parameter Mixture-of-Experts model that exceeds 1,000 tokens per second on a single 8-GPU commodity node. No custom silicon. The technical approach: FP4 quantization on MoE experts, DFlash speculative decoding, and TileRT's persistent kernel engine. Here is the technical review.