샤오미, 1T 모델 1000TPS 달성 'MiMo-V2.5-Pro-UltraSpeed' 공개

Blog Join us English 简体中文 Blog Join us English 简体中文 June 8, 2026 MiMo-V2.5-Pro-UltraSpeed: Pushing 1T-Parameter Model Generation Speed to 1000 TPS 1. Xiaomi MiMo-V2.5-Pro-UltraSpeed: Speed is the Ultimate Edge From the first roaring racer of the combustion age to the sonic boom that shattered the sound barrier, humanity's hunger for speed is written into our very DNA. The speed of AI reasoning is no different — it defines the boundaries of intelligence itself. When a model is fast enough, it ceases to be a tool you wait on and becomes an extension of your own thinking: responding in real time, iterating in an instant, collaborating without friction. Today, we are thrilled to release Xiaomi MiMo-V2.5-Pro-UltraSpeed in collaboration with TileRT, breaking the 1000 tokens/s decode speed on a 1-trillion-parameter model for the first time! 2. Limited-Time Access · Application-Based The MiMo-V2.5-Pro-UltraSpeed API launches simultaneously at a limited-time promotional price — 3× the cost of MiMo-V2.5-Pro, but delivering approximately 10× the generation speed! 3× the price, 10× the output experience. (API only; Token Plan not supported.) Due to limited high-speed inference resources, MiMo-V2.5-Pro-UltraSpeed will be available through an application-based, limited-time window. Approved users can access the API during the trial period, available only from June 9 to June 23, 2026, 23:59 (Beijing Time, UTC+8 / 08:59 PDT) . How to Apply API platform: platform.xiaomimimo.com/ultraspeed . Trial slots are limited — submission does not guarantee approval. We will prioritize enterprises and professional developers with genuine business needs. For standard model access, please follow the MiMo-V2.5 model series. For in-depth business partnerships for the UltraSpeed model, contact business-mimo@xiaomi.com . Chat Experience (Free During Trial) Approved users will receive free Chat access valid within the two-week window. Entry point: ultraspeed.xiaomimimo.com To ensure quality and fairness under resource constraints, the following rules apply: each account may enter the queue up to 10 times per day; each session is capped at 30 minutes; sessions idle for more than 5 minutes will be automatically released. 3. 1000 tokens/s: Not Just Fast, But a Paradigm Shift At the trillion-parameter (1T) scale, breaking 1000 tps is far more than a faster typewriter — it fundamentally disrupts AI application paradigms. First, speed itself begins to transmute into intelligence. Previously, when facing a hard problem, you could only "wait for one answer and pray it's correct." Now, within the same wall-clock time, the model can run dozens of reasoning paths in parallel (Best-of-N / Tree Search), automatically verifying and self-correcting in the background — using raw speed to generate depth of thought, directly elevating reasoning quality. Second, it completely unleashes the productivity ceiling of Coding Agents. Before, having AI write code meant developers painfully waiting in front of screens, bottlenecked by inference latency. At 1000 tps, code generation speed and production efficiency undergo a paradigm-level acceleration. Most importantly, trillion-parameter models can now enter real-time decision loops. Millisecond-level "think-respond" cycles allow 1T flagship models to seamlessly plug into time-critical scenarios — high-frequency quantitative trading signal generation, instant anti-fraud interception, intelligent bidding, and real-time interactive dialogue. And when this power is brought to surgical assistance and medical imaging analysis in life-or-death situations, AI speed is no longer just a metric of efficiency — it becomes a chip in the race against death. On the operating table, every second AI saves in completing lesion analysis and risk prediction gives the surgeon one more degree of freedom. This deepens our conviction that the ultimate significance of speed is not merely boosting productivity, but enabling technology to help humanity live better. 4. Extreme Model-System Codesign Achieving 1000+ tokens/s generation speed with a 1T flagship model is not the breakthrough of a single technique — it is the product of deep collaboration and extreme Codesign between the MiMo model team and the TileRT system team. The industry's current approach to similar extreme speeds typically relies on specialized hardware — Cerebras's Wafer-Scale integration or Groq's pure on-chip SRAM custom architecture. We chose a different path: achieving even more impressive inference speed on commodity GPUs through model-system codesign alone. On the model side, we applied FP4 quantization targeting the bandwidth bottleneck of commodity hardware, dramatically shrinking model size and reducing memory-access overhead; simultaneously, we introduced DFlash, an efficient speculative decoding method based on block-level masked parallel prediction , substantially increasing the accepted token length per verification step. On the system side, TileRT perfectly adapts to the dynamic characteristics of these algorithms, delivering a tailor-made compilation engine and compute kernels optimized specifically for the novel quantization and speculative decoding pipeline. Through this extreme Codesign, we achieved 1000+ tokens/s output from a 1T model using just a single standard 8-GPU commodity node. 3.1 FP4 Quantization At the trillion-parameter (1T) scale, traditional 8-bit (FP8 / INT8) or even 16-bit inference imposes prohibitive memory footprint and bandwidth pressure. Reducing parameter bit-width directly contributes to decoding speed. We therefore adopt the widely validated, virtually lossless FP4 (MXFP4) quantization format [1] . However, naively applying FP4 across the entire model causes degradation in complex reasoning, logic, and code generation. Given the MoE (Mixture of Experts) architecture of Xiaomi MiMo-V2.5-Pro — where Experts constitute the vast majority of parameters and exhibit the highest tolerance to quantization — we selectively quantize only the MoE Experts to FP4 while preserving original precision for all other modules. Through FP4 QAT (Quantization-Aware Training), we dramatically reduce model size and maximize hardware bandwidth utilization while keeping the model's overall capability essentially on par with the original, as shown below: 3.2 DFlash Speculative Decoding Traditional Speculative Decoding relies on a small draft model to "guess" subsequent tokens, which the large model then verifies. This transforms autoregressive generation (1 token per forward pass) into parallel multi-token generation, with rejection sampling during verification ensuring lossless output quality. However, its bottleneck lies in the draft model's quality determining the acceptance rate, while a stronger draft model incurs higher compute overhead — a fundamental tension. To break this deadlock, we adopt DFlash , an innovative block-level masked parallel prediction method from the research community [2] : the draft model fills an entire block of masked positions in a single forward pass, fundamentally eliminating the serial constraint of "autoregressive drafting." We deployed this approach on MiMo-V2.5-Pro with custom optimizations tailored for trillion-scale MoE and long-context scenarios. Using the Muon second-order optimizer and model self-distillation, we ensure that compact mask blocks still deliver ideal acceptance rates while compressing draft-stage overhead to near its theoretical minimum: The draft model exclusively uses Sliding Window Attention (SWA), naturally aligning with the SWA design of the MiMo-V2 series. This eliminates dependency on complete prefixes, reducing per-prediction compute from context-length-linear to constant. During training, mask-signal sampling is pushed down to GPU-local shards, enabling a single sequence to produce tens of thousands of independent training signals covering diverse context positions in one step — aligning with the long-c