Xiaomi MiMo-V2.5-Pro-UltraSpeed Release: 1T Model at 1000 Tokens Per Second

Xiaomi has announced the release of MiMo-V2.5-Pro-UltraSpeed, a 1-trillion-parameter (1T) model capable of breaking the 1000 tokens per second (tps) decode speed barrier. This achievement is particularly significant because it is realized on a single standard 8-GPU commodity node, rather than relying on specialized hardware like Wafer-Scale integration or custom SRAM architectures.

The Impact of 1000 TPS on AI Paradigms

Generating 1000 tokens per second transforms a trillion-parameter model from a latent tool into a real-time collaborator. This speed enables several paradigm shifts in AI application:

  • Speed as Intelligence: High throughput allows models to execute dozens of reasoning paths in parallel (via Best-of-N or Tree Search) and self-correct in the background within the same wall-clock time previously required for a single response.
  • Coding Agent Acceleration: The latency bottleneck for AI-driven code generation is effectively removed, allowing for near-instantaneous production efficiency.
  • Real-Time Decision Loops: Millisecond-level response cycles enable 1T models to be integrated into time-critical environments, such as high-frequency quantitative trading, instant anti-fraud systems, and real-time surgical assistance.

Technical Architecture: Model-System Codesign

Xiaomi achieved these speeds through a deep "codesign" between the MiMo model team and the TileRT system team, focusing on three primary technical pillars.

1. Selective FP4 Quantization

To address the memory bandwidth bottlenecks of commodity GPUs, Xiaomi implemented FP4 (MXFP4) quantization. To prevent the degradation of complex reasoning and logic typically associated with 4-bit quantization, the team applied a selective strategy:

  • MoE Experts: Since the Experts in the Mixture of Experts (MoE) architecture constitute the majority of parameters and are more tolerant to quantization, they are quantized to FP4.
  • Core Modules: All other modules preserve their original precision.
  • Optimization: Quantization-Aware Training (QAT) was used to ensure model capabilities remain essentially on par with the original FP8 version.

2. DFlash Speculative Decoding

Traditional speculative decoding is often bottlenecked by the serial nature of the draft model. Xiaomi utilizes DFlash, a block-level masked parallel prediction method:

  • Parallel Drafting: The draft model fills an entire block of masked positions in a single forward pass, eliminating autoregressive drafting constraints.
  • Efficiency: The draft model uses Sliding Window Attention (SWA), reducing compute from context-length-linear to constant.
  • Performance: In coding scenarios, the system achieves an average acceptance length of 6.30 tokens per verification round (with peaks of 7.14 out of 8), meaning the large model accepts the vast majority of the draft tokens.

3. TileRT Ultra-Low-Latency System

The TileRT inference system eliminates "execution gaps" caused by traditional operator boundaries through two primary innovations:

  • Persistent Engine Kernel: The entire compute pipeline remains resident on the GPU, allowing continuous prefetching of data while Tensor Cores are computing.
  • Warp Specialization: Communication, data movement, and tensor computation are decomposed into independent, coordinated thread groups (Warps), transforming the GPU into a heterogeneous execution system.

Availability and Pricing

MiMo-V2.5-Pro-UltraSpeed is available via an application-based trial from June 9 to June 23, 2026.

  • Pricing: The API is priced at 3× the cost of the standard MiMo-V2.5-Pro, while delivering approximately 10× the generation speed.
  • Open Source: The MiMo-V2.5-Pro-FP4-DFlash checkpoint, including FP4 quantized weights and DFlash parameters, has been open-sourced on HuggingFace.

Community Insights and Perspectives

Industry observers and developers have noted several implications of this release:

  • Competitive Landscape: Some users suggest that the aggressive pricing and speed optimizations from Chinese providers may pressure American AI companies. One user noted:

    "These price and speed optimization from Chinese providers, combined with the raising prices from American ones will change the game sooner than later."

  • Developer Workflow: The shift toward near-instant responses is seen as a potential game-changer for "human-in-the-loop" coding, moving away from "vibecoding" toward a more partner-like interaction.
  • Skepticism on Utility: Not all users view speed as the primary bottleneck. Some argue that for regular employees, the ability to complete tasks faster does not necessarily reduce the 8-hour workday, but rather changes the nature of the work to a "slot machine" of prompting.
  • Technical Critique: Some developers questioned if the innovations (such as persistent kernels and warp specialization) are truly novel or simply standard CUDA optimizations applied at scale.

Sources