Unveiling Hunter Alpha: Xiaomi’s MiMo‑V2‑Pro and Two New Models Revealed

A week before the official announcement, two unnamed models—Hunter Alpha and Healer Alpha—appeared on OpenRouter, repeatedly topped the daily leaderboard and accumulated over 1 trillion token calls, prompting worldwide speculation about their origin.

1. Identity Reveal

On the night of March 19 2026, Xiaomi confirmed that Hunter Alpha corresponds to MiMo‑V2‑Pro (a flagship agent model) and Healer Alpha to MiMo‑V2‑Omni (a multimodal base model), and simultaneously announced the speech‑synthesis flagship MiMo‑V2‑TTS.

2. MiMo‑V2‑Pro: Flagship Agent

MiMo‑V2‑Pro is designed for the emerging Agent era with the following core specifications:

Total parameters: over 1 trillion (42 billion active parameters)

Architecture: Hybrid Attention

Context window: up to 1 million tokens (native long‑context support)

Global AI Index rank: 8th worldwide, 2nd in China

Benchmark results highlight its competitiveness:

PinchBench: 84.0 (3rd globally)

ClawEval: 61.5 (3rd globally, close to Claude Opus 4.6)

SWE‑bench Verified: 78.0

Terminal‑Bench 2.0: 57.1

The official positioning states that its coding ability surpasses Claude Sonnet 4.6, overall Agent performance approaches Claude Opus 4.6, and its API pricing is only one‑fifth of Opus. In Artificial Analysis’s “high cost‑performance model ranking” (≤ $0.15 per M tokens), MiMo‑V2‑Pro scores 49 points, ranking first and handling 500 billion tokens per week—the platform’s most used model. API pricing (within 256 K context) is $1 per M input tokens and $3 per M output tokens.

3. MiMo‑V2‑Omni: Full‑Modal Agent

MiMo‑V2‑Omni extends the Agent stack with vision, audio, video, and cross‑modal reasoning capabilities:

Audio understanding: exceeds Gemini 3 Pro, supports continuous audio comprehension for over 10 hours, capable of environmental sound classification and speaker separation

Visual reasoning: multi‑disciplinary visual inference and complex chart analysis, surpasses Claude Opus 4.6 and approaches Gemini 3 Pro

Video understanding: native audio‑video joint input, situational awareness, future prediction, and a novel video pre‑training architecture

Agent abilities: cross‑modal perception, autonomous planning, real‑time strategy correction, end‑to‑end result delivery

It offers a 256 K token context window and API pricing of $0.4 per M input tokens and $2 per M output tokens.

4. MiMo‑V2‑TTS: Speech Synthesis Flagship

MiMo‑V2‑TTS is Xiaomi’s self‑developed large‑scale speech synthesis model, built on a proprietary Audio Tokenizer and a multi‑codebook speech‑text joint modeling architecture trained on tens of billions of hours of audio data. Multi‑dimensional reinforcement learning enables fine‑grained style control.

Multi‑granular style control: from overall tone to local emotion, adjustable with precision

Dialect support: Northeast, Sichuan, Henan, Cantonese, Taiwanese accents

Singing capability: accurate pitch and rhythm expression

Emotion switching: sentence‑level tone changes with natural emotional transitions

Voice cloning: high‑fidelity speaker replication

This signals Xiaomi’s intention to integrate speech as an essential layer of the AI Agent capability stack.

5. Blind‑Test Marketing Strategy

Instead of a traditional pre‑launch hype, Xiaomi employed a “blind test” approach: the models ran anonymously on OpenRouter for a full week, gathering real user data and reputation before the brand reveal. This avoided preconceived bias toward domestic models and prevented credibility loss from overt self‑promotion.

Hunter Alpha topped OpenRouter’s daily leaderboard for multiple days.

Total token usage exceeded 1 T, making it the platform’s most invoked model.

Developers worldwide speculated about the model’s origin, sparking discussions on communities such as LINUX DO and X.

Community feedback highlighted the surprise factor, with users initially assuming the models were from major closed‑source providers before learning they were Xiaomi’s.

6. Product and API Availability

All three models are available immediately, covering both product‑side and developer‑side integration:

Product side: Xiaomi Miclaw, MiMo Studio, Kingsoft Office, Xiaomi Browser

Agent frameworks: OpenClaw, OpenCode, KiloCode, Blackbox, Cline

API access: platform.xiaomimimo.com (free trial during the launch week)

The comprehensive Agent framework coverage ensures low switching costs for developers, allowing testing within familiar tools.

7. Editorial Assessment

The anonymous performance validation demonstrates genuine technical strength—Hunter Alpha achieved top rankings and 1 T token usage without brand backing, reflecting authentic developer endorsement rather than artificial ranking manipulation.

The three‑model matrix—Pro (flagship inference), Omni (full‑modal perception), TTS (speech synthesis)—covers the core Agent capability stack of perception, reasoning, and expression in a single release.

Cost‑performance is striking: comparable Agent performance to Claude Opus 4.6 at one‑fifth the price, making the suite highly attractive for enterprises planning large‑scale AI deployment.

Overall, Xiaomi’s MiMo‑V2 series adds a serious contender to the competitive landscape of large‑scale domestic AI models.