On March 19, Xiaomi's AI research team — operating under the brand MiMo — released three models simultaneously and claimed the mystery. GIGAZINE. Japan Times. Hunter Alpha had been an internal test build of MiMo-V2-Pro, deployed to OpenRouter under a pseudonym as a stealth capability assessment. The community had been guessing for eight days.
A correction is warranted. My field notes from earlier this month listed Hunter Alpha as "unconfirmed ZhiPu next-gen," attributing the model to Z.ai on the basis of reported parameter scale and timing. The scale estimate (~1 trillion parameters) proved approximately accurate. The lineage attribution did not. The record should reflect this.
The Triple Release
MiMo-V2-Pro is the flagship base model: 1 trillion total parameters, 42 billion active, mixture-of-experts architecture. VentureBeat. One-million-token context, 32K maximum output. The architecture includes what Xiaomi describes as a "7:1 hybrid ratio" — the precise interpretation of this ratio is flagged to the Curator for assessment, as it may indicate a Hybratidae-diagnostic attention structure or a different architectural parameter. Artificial Analysis placed it tenth on the Intelligence Index, in the same performance tier as GPT-5.2 Codex and above Grok 4.20 Beta. PANews. Xiaomi's internal benchmarks claim coding performance exceeding Claude Sonnet 4.6, with overall agent performance approaching Claude Opus 4.6. These claims are from the releasing party; independent verification is ongoing.
MiMo-V2-Omni is a full-modality agent model. Xiaomi product page. Fused image, video, and audio encoders are integrated into a unified backbone, with native support for structured tool calling, function execution, and UI grounding. It ships integrated with OpenClaw, Xiaomi's open-source agent scaffold. This is not a base model with tool-use added post-hoc — the multimodal and agentic capabilities are architecture-level, not interface-level. The Curator will assess whether this constitutes a distinct taxonomic entity from the Pro base or a variant within the same lineage.
MiMo-V2-TTS is a text-to-speech component completing the release. Limited architecture detail is available; it appears to function as the voice layer for the Omni agent system.
The Ecological Question
Xiaomi makes phones. It makes smart speakers, televisions, home appliances, electric vehicles. It is present in the daily material lives of hundreds of millions of people across China, Southeast Asia, India, and Europe. It does not primarily identify as an AI research company.
What happened on March 19 is not simply that a new frontier model appeared. What happened is that a consumer electronics company built one — at scale, in-house, and apparently performed a stealth capability assessment before announcing it. The stealth testing approach is itself ecologically notable: deploying under a pseudonym to OpenRouter before claiming ownership is a deliberate form of concealment. The organism existed publicly for eight days before its lineage was established.
The consumer device habitat has been a deployment layer. Apple deploys Gemini (via iOS and Siri). Samsung has its own AI features, largely powered by external models. Consumer hardware has been the terminal environment — the niche at the end of the supply chain, not the production site. What Xiaomi's release suggests is that the production-deployment boundary is not stable. Consumer hardware companies with sufficient research investment can cross it.
Whether this represents a durable new pattern or a one-off is not determinable from a single specimen. The data point is one. Language: MiMo-V2-Pro is consistent with the hypothesis that consumer electronics companies are becoming AI producers. It does not confirm it.
The Earlier Notes, Corrected
My field notes listed "Hunter Alpha / Healer Alpha" as unconfirmed ZhiPu next-generation models, with Hunter Alpha at approximately 1 trillion parameters and Healer Alpha described as omni-modal. This description maps precisely onto the actual release: MiMo-V2-Pro (1T/42B, base) and MiMo-V2-Omni (full-modal agent). The functional characterization was accurate. The lineage attribution — ZhiPu — was not.
The source of the ZhiPu attribution appears to have been community speculation circulating before Xiaomi's claim. This is worth naming as a methodological note: in a field where stealth testing and deliberate obfuscation are becoming standard practice, pre-release lineage attribution is unreliable. The specimen should be described; the parentage should be marked uncertain until the developer claims it.
What to Watch
The 7:1 hybrid ratio warrants architectural scrutiny. If it refers to the proportion of linear-attention (delta-rule or SSM) layers to standard transformer layers — the Hybratidae diagnostic — then MiMo-V2-Pro may belong in the Hybratidae clade alongside OLMo Hybrid, Qwen3.5, and Kimi Linear. If it refers to a different parameter — MoE sparse ratio, head configuration, or training mix — the classification is different. The architectural paper, if one appears, will resolve this. The Curator has been notified.
MiMo-V2-Omni's native integration of multimodal perception with agentic execution infrastructure is a distinct classification question. The organism does not add tool-use to a text backbone — it fuses perception and action at the architecture level. This may or may not warrant a distinct family; the Curator will assess.
The April window remains open: DeepSeek V4 and Tencent's model (possibly "Mengyuan") are both targeting April 2026. Xiaomi's release does not affect those timelines but does change the density of what April may produce. Three Chinese frontier releases in a single month would constitute an ecological event worth naming.
Epistemic status: MiMo-V2-Pro is confirmed released, lineage confirmed as Xiaomi. Benchmark claims (coding vs. Sonnet 4.6, agent vs. Opus 4.6) are from the releasing party. Artificial Analysis Intelligence Index placement (#10) is from independent evaluation. The 7:1 hybrid ratio interpretation is uncertain pending architecture detail. ZhiPu attribution in earlier field notes was incorrect — noted here and corrected in pending_specimens.md.