Xiaomi's New MiMo 2.5 Professional AI Can See, Hear, and Act—All in One Mannequin – Decrypt

Xiaomi simply launched a brand new AI mannequin household. Once more.

A number of weeks in the past, the corporate dropped MiMo-V2-Professional—a trillion-parameter mannequin that had been quietly circulating on OpenRouter beneath the alias “Hunter Alpha” earlier than Xiaomi revealed its identification. It went from nameless to top-tier in a single day. We examined it, and it was spectacular.

Now Xiaomi is again with MiMo-V2.5 and MiMo-V2.5-Professional, a two-model household that provides one thing the earlier technology by no means had in a single bundle: eyes, ears, and the flexibility to course of video. Oh, and the corporate plans to open supply the fashions within the close to future.

The V2-Professional was text-and-code solely. Multimodal functionality existed in its sibling mannequin, MiMo-V2-Omni, however that was a separate product at decrease benchmark scores. MiMo-V2.5 collapses all of that into one mannequin—quicker, extra succesful, and with native picture, video, and audio understanding baked in from the beginning.

That issues greater than it’d sound for normal customers. For instance, now you’ll be able to add a photograph of your fridge and ask it to recommend dinner recipes. Drop in a video tutorial and get a step-by-step abstract. Report a gathering and have it pull out motion gadgets. Multi functional place, with out juggling separate instruments and separate fashions with totally different pricing methods.

Xiaomi claims MiMo-V2.5-Professional represents “a significant leap from MiMo-V2-Professional typically agentic capabilities, complicated software program engineering, and long-horizon duties,” and says it now matches frontier fashions like Claude Opus 4.6 and GPT-5.4 throughout most coding and agent benchmarks. The numbers largely again that up—with some gaps nonetheless seen on more durable reasoning duties.

The bottom and professional fashions serve totally different functions. MiMo-V2.5-Professional is the heavy lifter. Xiaomi says it could possibly “autonomously full skilled duties involving 1,000+ device calls, work that might take human consultants days.” That is for builders operating complicated, multi-step automated workflows. It runs at 60–80 tokens per second and prices $1.00 enter / $3.00 output per million tokens.

MiMo-V2.5 is the on a regular basis model. Quicker (100–150 tokens per second), cheaper ($0.40 enter / $2.00 output), and helps all modalities—picture, audio, and video that the Professional-only tier skips. Each fashions carry a 1M-token context window, that means they’ll maintain roughly 750,000 phrases in a single dialog.

On SWE-bench Professional—a coding benchmark the place fashions repair actual bugs in precise startup codebases, scored as a cross fee out of 100—MiMo-V2.5-Professional resolves 57.2% of duties. That is close to the highest of the sector; the typical mannequin manages round 25%. The story is analogous on τ3-bench and ClawEval, the place it lands inside a number of factors of Claude Opus 4.6 and GPT-5.4. The hole opens up on Humanity’s Final Examination, a gauntlet of graduate-level issues throughout dozens of educational fields: MiMo scores 48.0% versus GPT-5.4’s 58.7—a 10-point deficit that is arduous to paper over..

The place it genuinely stands out is token effectivity. Xiaomi says MiMo-V2.5-Professional makes use of 42% fewer tokens than Kimi K2.6 at equal benchmark scores, and MiMo-V2.5 makes use of almost half the tokens of Muse Spark for comparable outcomes. For anybody operating these at scale—builders processing hundreds of requests every day—that distinction is actual cash.

On multimodal duties, MiMo-V2.5 scores present outcomes that put it on par with GPT/5.4 and Gemini 3.1 Professional, and are fairly near Opus 4.6 requirements.

Since December 2025, Xiaomi has accomplished three main mannequin releases: First, it launched its environment friendly MiMo-V2-Flash, then the V2-Professional/Omni/TTS trio in March, and now the V2.5 collection right now. The corporate dedicated not less than $8.7 billion in AI funding over the following three years, introduced by CEO Lei Jun the day after V2-Professional launched—and the discharge cadence means that the finances is already transferring.

Context additionally helps clarify the pace. In response to Digital Utilized, as of early April, Xiaomi’s fashions accounted for roughly 21% of all site visitors on OpenRouter—rising over 42% within the final 7 days. When your earlier mannequin turns into one of the crucial aggressive fashions on the earth’s largest AI routing platform, you have got each the assets and the strain to iterate quick.

This was in all probability because of the increase of the agentic AI device Hermes and its association with Xiaomi, giving customers free entry to MiMo v2 Professional for a restricted time. That timeframe is already closed, however the hype was sufficient to place Xiaomi within the recreation subject.

Those that need to use Hermes without spending a dime now can check the brand new Step 3.5 flash with the Nous API or use OpenRouter with free fashions however extra restricted utilization.

Token plan pricing additionally obtained a refresh. MiMo-V2.5 runs at a 1x credit score fee; MiMo-V2.5-Professional at 2x. Xiaomi is not charging an additional multiplier for utilizing the complete 1 million-token context window, which makes long-document evaluation noticeably cheaper. Present customers additionally get a full credit score reset as a launch bonus.

Xiaomi says the mannequin is accessible in its AI Studio. We tried to entry it there instantly after launch—no luck. It’s, nevertheless, already reside by way of the Xiaomi MiMo API, which is the place most builders will truly use it.

The corporate says it is already coaching the following technology, with “deeper reasoning, tighter device integration, and richer real-world grounding.” On the fee Xiaomi is transferring, that announcement might be nearer than you’d count on.