MiMo-V2.5-Pro-UltraSpeed Hits 1000 Tokens Per Second: The Operator's Routing Decision

Speed rarely changes operator routing decisions on its own. What changes decisions is when speed alters the economics of an entire workload class. MiMo-V2.5-Pro-UltraSpeed — Xiaomi and TileRT's new inference mode — crosses exactly that threshold: 1000+ tokens per second from a 1-trillion-parameter model, delivered over a standard OpenAI-compatible API. The trial window opens today, June 9, and closes June 23.

That is the operational fact. The harder question is whether your workloads justify the 3× price premium. This article is a decision framework, not a product endorsement.

What Happened: MiMo-V2.5-Pro-UltraSpeed 1000 Tokens Per Second

On June 8, 2026, Xiaomi released MiMo-V2.5-Pro-UltraSpeed in collaboration with TileRT, achieving 1000+ tokens/s decode speed from a 1T-parameter model — a claimed first on commodity GPU hardware. The API model ID is mimo-v2.5-pro-ultraspeed, served at https://api.xiaomimimo.com/v1, fully OpenAI-compatible.

The technical approach is notable precisely because it avoids custom silicon. The two key techniques:

• FP4 quantization (MXFP4) on MoE Experts only. Because MiMo-V2.5-Pro is a Mixture-of-Experts architecture, the Expert layers (the vast majority of parameters) are quantized to FP4 via QAT (Quantization-Aware Training), while attention and non-Expert modules retain original precision. The result is dramatically lower memory bandwidth pressure — the core bottleneck at 1T scale — with benchmark performance described as "essentially on par with the original."

• DFlash speculative decoding. Rather than a separate draft model, DFlash uses block-level masked parallel prediction to increase the number of accepted tokens per verification step. This transforms the single-token autoregressive bottleneck into parallelized multi-token verification, preserving output quality while multiplying throughput.

The API is application-only during the trial period (June 9–23, 2026, 23:59 Beijing Time). Slots are limited; trial slots do not guarantee approval. Standard mimo-v2.5-pro remains available without restriction.

Pricing signal: UltraSpeed costs 3× the base MiMo-V2.5-Pro price while delivering approximately 10× the generation speed. Token Plan billing is not supported; API-only.

Why It Matters for AI Engineering Teams

The 1000 TPS threshold is not just a benchmark. At this speed, several workload patterns change character:

Coding agents with tight inner loops. Agentic coding workflows that generate, execute, and re-plan in short cycles are gated on model response latency. At 1000 TPS, the model's share of wall-clock time collapses from the dominant factor to near-negligible. If your Claude Code or Cursor backend completes a code suggestion in 1.2 seconds instead of 12 seconds, the development feedback loop changes qualitatively — not just quantitatively.

Best-of-N and tree search reasoning. With sufficient throughput, you can afford to run N parallel reasoning paths within the same latency budget you previously spent on a single path. At 1000 TPS, best-of-N sampling over a 1T reasoning model becomes economically viable on per-call budgets that previously justified only a single pass.

Real-time decision pipelines. The release notes high-frequency trading signal generation, anti-fraud, and intelligent bidding as target scenarios. For operators building these: the key variable is not TPS in isolation, but TPS relative to your tolerable first-token latency. UltraSpeed reduces generation time; TTFT depends on system load during the limited trial.

The Router/Operator Angle: When Does 3× Price Justify 10× Inference Speed for Coding Agents?

The honest routing analysis starts with unit economics. Here is a decision framework for operators evaluating mimo-v2.5-pro-ultraspeed against the standard endpoint:

Workload type → routing candidate?

The 3×/10× math. If your average output is N tokens, UltraSpeed costs 3× per token and delivers 10× the tokens per second. The cost-per-output is 3× higher. But the wall-clock time to complete a fixed-token task is ~10× lower. For workflows where developer wait time is the binding constraint — not token spend — UltraSpeed's effective ROI is positive even at 3× price.

OpenAI-compatible integration. The model follows the standard chat completions API with streaming. Switching from mimo-v2.5-pro requires only a model name change. For operators already routing through an AI gateway with multiple Chinese providers, this is a zero-code test — add the model ID to your routing table, point traffic at it for a canary segment, and measure.

Access risk during trial. Trial access is application-only, resource-constrained, and time-bounded to June 9–23. Operators evaluating UltraSpeed should:

1. Apply immediately at platform.xiaomimimo.com/ultraspeed if you qualify as a business or professional developer.
2. Treat this as a benchmark window, not a production dependency.
3. Define a fallback route to mimo-v2.5-pro or another provider for workloads that exceed trial quotas.

What TheRouter Users Should Watch or Try

For teams routing Chinese provider models through an AI gateway, the UltraSpeed trial is worth instrumenting even if you do not immediately commit to the premium tier. The data to capture:

• TTFT under load during trial hours — this is the missing variable in the "10× speed" claim.
• Benchmark your specific coding agent loop — token counts and round-trip latency for your actual SWE-bench-equivalent inner loop, not a synthetic throughput test.
• Cost-per-accepted-code-edit as the unit of comparison, not cost-per-million-tokens.

The provider routing guide on TheRouter covers the pattern for adding and comparing Chinese provider model endpoints. For multi-provider fallback configuration when trial quotas are hit, see the AI provider fallback routing docs.

When the trial closes June 23, the production availability and pricing of mimo-v2.5-pro-ultraspeed will determine whether this becomes a routing-table permanent resident or a benchmark curiosity. Watch the Xiaomi MiMo announcements for GA pricing.