Xiaomi Says It Runs a 1-Trillion-Parameter Model at 1,000 Tokens a Second. Should You Believe It?

A phone maker just claimed it built a one-trillion-parameter AI model. Fine, plenty of companies are AI companies now. But then it claimed that model spits out 1,000 tokens per second. That combination is where it gets interesting. Xiaomi’s MiMo-v2.5-Pro-UltraSpeed has been quietly making the rounds in tech circles, and on paper the numbers are eye-catching. Before you take them at face value, though, there’s a lot worth interrogating.

Let me be upfront about something. The verified material on this is thin. Over the past month there’s been almost no substantive discussion on Reddit or X, and the trail more or less dead-ends at a single YouTube video. So this isn’t a “here are the confirmed facts” piece. It’s a “here’s what these numbers would mean, and how far you should trust them” piece.

Why “huge” and “fast” don’t usually share a sentence

Start with the fact that both numbers appear at once. Big models are normally slow. The more parameters you have, the more math each token requires. One trillion (1T) parameters puts you in the same weight class as the largest frontier models on the planet.

But 1,000 tokens per second is the kind of speed you usually see from small, nimble models. For context, a fast human reader manages a few hundred words per minute. A thousand tokens per second is text pouring out at dozens of times the speed you can read it. So this single model is claiming to be both enormous and blistering — and technically, that’s not a combination you just wave through.

The trick has a name: MoE

The likely key to this contradiction is Mixture of Experts (MoE). The fact that Xiaomi bolted “UltraSpeed” onto the name reads like a deliberate nod to exactly this.

Here’s the plain-English version. Instead of running all trillion parameters for every token, the model is carved up internally into many “experts.” When a query comes in, only a handful of those experts actually fire. Think of a 1,000-person company where any given task pulls in just a few people — not the whole payroll.

So the total parameter count can be a trillion while the number of parameters actually activated per token is far smaller. That gap is where the speed comes from. Which means if someone leans hard on “it’s a trillion-parameter model and it’s this fast,” that’s at best a half-truth. An honest comparison has to show the active parameter count alongside it.

Under what conditions was “1,000 tokens per second” measured?

Speed numbers always come with fine print. The same model can vary several-fold depending on the setup, so there are a few things you should always check on any marketing figure.

First, batching. If that throughput is the aggregate across many requests bundled together, it’s not the speed a single user actually feels.

Second, the hardware. Numbers pulled from a rig with several top-end accelerators bolted together look nothing like numbers from an ordinary server.

Third, dedicated inference chips. A lot of recent speed records owe more to specialized inference infrastructure than to the model itself. In other words, it may not be “the model is fast” so much as “it’s fast when you run it on this specific stack.”

A “1,000 tokens per second” with none of those three conditions disclosed is a number with enormous room for interpretation.

Why speed became the new battleground

Step back for the bigger picture. For a while, the AI race was about how smart a model was — benchmark scores, full stop. But as those scores converged near the top, the center of gravity has shifted toward speed and cost.

The logic is simple. Once a model is smart enough, the next question becomes “how cheaply and quickly can I run it?” Faster inference means more users per server, which means lower cost per query. For anything that demands instant responses — real-time voice assistants, agents — that math matters even more.

Xiaomi putting “UltraSpeed” front and center looks like a company reading that shift correctly. For a firm that needs to bake AI into everything from phones to appliances to EVs, “fast and cheap inference” isn’t an abstract brag. It’s closer to a survival requirement.

So how much should you believe?

Cold sober take: right now there’s almost no independently verified benchmark and no credible third-party testing. As noted, community discussion is sparse and the sourcing is limited. At this stage, “one trillion parameters, 1,000 tokens per second” is safest treated as a claim, not a verified fact.

That doesn’t make it noise, either. The broader pattern — Chinese big tech pushing aggressively not just on model size but on inference speed and cost — is real and unmistakable. Whether Xiaomi’s specific announcement is genuine or inflated, it’s a solid signal that the contest is migrating from “scores” to “speed and efficiency.”

The checklist is clear. How many parameters are actually active? On what hardware and under what conditions was the speed measured? And can a third party reproduce the same result? Until those three things are public, the wise move is to hang a question mark on the number rather than get excited by it.

What’s your read? Are we entering an era where “how fast and how cheap” matters more than “how smart”?