MiMo-V2.5-Pro-UltraSpeed, launched jointly by Xiaomi's MiMo team and the AI inference systems team TileRT, is a counterintuitive model. It puts a trillion-parameter (1T) MoE model on a single standard 8-GPU commodity node, breaks 1000 tokens/s of output for the first time, and peaks around 1200 tokens/s — without relying on dedicated hardware like Cerebras wafer-scale chips or Groq's custom SRAM. Within two weeks of launch it received more than 66,000 usage applications, forcing an extension of the trial window originally set to end on June 23. This article breaks down its technical approach, pricing, and how to apply, so you can judge where the speed actually comes from and whether it's worth using.

What Is MiMo-V2.5-Pro-UltraSpeed

One-line definition: Xiaomi's trillion-parameter flagship model running at 1000 tokens/s on commodity hardware.

• Architecture: MoE (Mixture of Experts)
• Total parameters: 1T (one trillion)
• Activated parameters: about 42 billion per forward pass
• Context length: supports 1M tokens
• Core breakthrough: stable 1000 tokens/s output on an 8-GPU commodity node, peaking around 1200 tokens/s

Xiaomi's MiMo announcement said applications had "far exceeded expectations," and the open window would be extended.

Speed: Where It Sits in the Industry

Placing 1000 tokens/s on the industry map makes the impact obvious. According to the AI benchmarking platform Artificial Analysis:

That puts UltraSpeed's output speed an order of magnitude faster than frontier flagship models. For latency-sensitive scenarios — real-time conversation, streaming code generation, bulk long-form production — the difference in experience is a qualitative leap.

How It Does It: Joint Model-Side and System-Side Optimization

The key is that it relies on no dedicated hardware, instead squeezing performance out of commodity GPUs through joint optimization on the model side and the system side.

Model Side

• FP4 mixed quantization: FP4 is applied mainly to the MoE experts while other modules keep higher precision, cutting model size and memory-traffic pressure
• DFlash speculative decoding: block-wise masked parallel prediction replaces the token-by-token autoregression of a traditional draft model, letting the large model verify more candidate tokens at once

The open-sourced Hugging Face model card for MiMo-V2.5-Pro-FP4-DFlash is the base model behind UltraSpeed: an FP4-quantized backbone plus a BF16 DFlash drafter, under the MIT license.

System Side

• TileRT built a custom compilation engine and compute kernels for the FP4 quantization and DFlash pipeline
• A resident kernel engine and heterogeneous pipeline cooperation reduce kernel-launch and synchronization overhead

Pricing: 3x the Price for 10x the Speed

The UltraSpeed API uses a limited-time trial price set at 3x the standard MiMo-V2.5-Pro, delivering roughly a 10x boost in output speed.

For reference, Anthropic's flagship Claude Opus is publicly priced at $5 per million input tokens (about 34 RMB) and $25 per million output tokens (about 170 RMB). UltraSpeed's output price is only about 1.5% of Claude Opus's, while being more than ten times faster — of course, model capability can't be equated one-to-one, but that price-performance ratio is very friendly to throughput-heavy workloads.

Lei Jun posted on Weibo to announce the new MiMo-V2.5-Pro-UltraSpeed milestone.

How to Apply for Access

The model is still in a limited-time trial: users already approved can keep using it, and new users need to apply for the beta:

• API application: https://platform.xiaomimimo.com/ultraspeed
• Chat trial: https://ultraspeed.xiaomimimo.com
• Open-sourced base model: https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash (MIT license)

As of June 23, more than 66,000 usage applications had come in, from Fortune Global 500 companies, industry leaders, and individual developers, across legal, finance, telecommunications, logistics, auto manufacturing, media, and higher education.

Use Cases

1000 tokens/s unlocks a batch of scenarios that previously "couldn't afford to wait":

• Real-time conversation and support: first-token and streaming latency near real time — a qualitative user-experience jump
• Streaming code generation: no more stutter in an coding agent's multi-round iterations
• Bulk long-form production: a major throughput lift for reports, marketing copy, and knowledge-base cleanup
• Multi-agent orchestration: agent-to-agent calls are latency-sensitive, and ultra-fast output cuts total wall time significantly
• Education and companionship: conversational, instant-feedback scenarios demand the fastest possible responses

Controversies and Limitations

• Comparability of "1T parameters": developers overseas have questioned how "trillion parameters" in a MoE architecture compares with dense models; the fact that only about 42 billion parameters activate needs to be spelled out
• Ecosystem maturity: compared with mature ecosystems like Claude and GPT, MiMo's tooling and agent-framework support are still catching up
• An objective gap in capability: speed advantages don't translate directly into reasoning advantages; for complex reasoning and long-chain agent tasks, run your own benchmarks through the Chat entry point first

💡 Tip: if your workload is throughput-heavy (support, bulk production, multi-agent orchestration), UltraSpeed's price-performance is highly attractive; if your core need is complex reasoning, run a few benchmarks through the Chat entry point before migrating.