When the anonymous "Hunter Alpha" and "Healer Alpha" models appeared on OpenRouter, the community spent a week guessing which lab was behind them. Xiaomi broke the suspense during its MiMo event: Hunter Alpha was the stealth test build of MiMo-V2-Pro, and Healer Alpha was MiMo-V2-Omni. With the mystery solved, we can now treat them as Xiaomi's flagship agent duo—one optimized for trillion-scale reasoning, the other for native multi-modal perception. This article distills what Xiaomi actually announced, why the MiMo pair matters, and how to integrate them without losing sight of privacy or governance.
TL;DR
- MiMo-V2-Pro (Hunter Alpha): Sparse trillion-parameter architecture with a full 1M-token context, tuned for agent planning, SWE-Pro workloads, and research-scale reasoning.
- MiMo-V2-Omni (Healer Alpha): Full-modality agent (text/image/audio/video) with 262K tokens of context and sub-100 tokens/second latency for perception-heavy execution.
- Testing route: Xiaomi pushed both models to OpenRouter under the Hunter/Healer codenames to collect unbiased usage telemetry before the public MiMo reveal.
- Cost and access: MiMo-V2 APIs are staying on OpenRouter for the near term while Xiaomi deploys them into Kingsoft WebOffice, Xiaomi Browser, and device-side copilots.
- Risk: Prompts and completions are logged by Xiaomi; treat the free tier as a public testbed until private tenancy is available.
What Xiaomi Announced
Xiaomi's MiMo line now has three pillars—V2-Pro (reasoning), V2-Omni (perception), and V2-TTS (expressive voice). The company's positioning is simple: "ompress the complex world into language, then project language back into physical action." That is why Hunter Alpha focused on sparse trillion-scale logic while Healer Alpha focused on fused modalities.
MiMo-V2-Pro (Hunter Alpha codename)
MiMo-V2-Pro is the “brain” of Xiaomi’s agent stack. The table summarizes the specs that map to the Hunter Alpha experience:
Attribute | Detail |
Architecture | Sparse MoE (1T parameters, 42B active) |
Context Window | 1,048,576 tokens |
Workloads | SWE-Pro, PinchBench, ClawEval, end-to-end DevOps copilots |
Latency & Cost | Comparable to GPT-5.2 latency, a fraction of Opus pricing per Xiaomi benchmarks |
Deployments | Xiaomi Browser, Kingsoft WebOffice, OpenRouter API ( |
Logging | Prompts/completions logged by Xiaomi for model tuning |
Your AI Receptionist, Live in Minutes.
Scale your front desk with an AI that never sleeps. Solvea handles unlimited multi-channel inquiries, books appointments into your calendar automatically, and ensures zero missed opportunities around the clock.
Why it matters
Teams chasing long-horizon reasoning can finally use a 1M-token window without paying GPT-5 prices. Xiaomi's internal tests showed MiMo-V2-Pro matching Claude Opus 4.6 on system design tasks while running at roughly half the cost. Because it already powers Xiaomi's own services, it ships with production-grade agent harnesses—workflow DAGs, retry logic, and telemetry hooks—that you can reuse inside OpenClaw or your own orchestrator.
MiMo-V2-Omni (Healer Alpha codename)
MiMo-V2-Omni serves as the native sensor layer. It can read dense spreadsheets, watch QA videos, listen to customer calls, and stream context to MiMo-V2-Pro.
Attribute | Detail |
Modalities | Text, image, audio, video (native fusion) |
Context Window | ~262,000 tokens |
Throughput | ~93 tokens per second (OpenRouter listing) |
Real-World Skills | UI draft + HTML handoff, field ops surveillance, multi-channel design reviews |
Deployments | Xiaomi Creator tools, Alpha Stack references, OpenRouter API ( |
Logging | Same as Pro—provider logs prompts and completions |
Why it matters
Healer Alpha turns agent workflows into multi-sensory experiences. Designers feed it screenshots and voice notes, and it returns component libraries plus Hunter-ready prompts. Field teams send it camera footage, and it returns normalized alerts for MiMo-V2-Pro to triage. Because Xiaomi uses Healer inside its own hardware ecosystem, the model already understands MIUI screens, IoT telemetry, and browser contexts out of the box.
Xiaomi's Alpha Stack Strategy
Layer | Role | Xiaomi Implementation |
Perception | Convert multimedia evidence into structured context | MiMo-V2-Omni (Healer Alpha) |
Cognition | Plan, reason, refactor, document | MiMo-V2-Pro (Hunter Alpha) |
Voice/Action | Speak, sing, deliver expressive responses | MiMo-V2-TTS |
Routing | Manage skills, guardrails, analytics | Xiaomi Agent runtime + OpenRouter sandboxes |
Takeaway: Xiaomi's own devices route perception → cognition → voice in one stack. External teams can mirror the same pipeline by pairing Healer (ingestion) with Hunter (planning) and finishing with your existing toolchain.
Real-World Scenarios
Early beta customers highlighted four high-impact scenarios:
- AI office suites: Healer ingests PPT/PDF/voice feedback, Hunter rewrites decks, and MiMo-V2-TTS pitches the output in natural Mandarin or English.
- Xiaomi Browser assistant: Healer interprets multi-tab sessions while Hunter summarizes and cross-checks sources for the user.
- Manufacturing QA: Healer watches conveyor footage; Hunter proposes remediation steps and links them to MES tickets.
- Developer copilot: Healer captures design critiques and log snippets; Hunter pushes full refactors and writes postmortems.
Risks and Governance
- Data residency: Xiaomi runs the models on its own infrastructure; confirm whether your compliance program allows data egress to Xiaomi-operated clouds.
- Logging: Free-tier usage is logged. Redact secrets or wait for enterprise tenancy before connecting regulated systems.
- Spec maturity: Xiaomi has not released open weights. Any regression or downtime is tied to their roadmap, not OpenRouter’s.
- Dependency clustering: Relying on both Hunter and Healer means betting on Xiaomi's entire stack—plan for fallback models in case pricing or access changes.
Integration Checklist
To adopt MiMo-V2 without surprises:
- Confirm licensing: Xiaomi's enterprise team can issue contracts; OpenRouter alone does not provide SLAs.
- Segment workloads: Route long-context reasoning to MiMo-V2-Pro, and route multimedia ingestion to MiMo-V2-Omni.
- Instrument logging: Mirror Xiaomi's logs inside your SIEM so you can reconcile what was sent to OpenRouter.
- Add fallback policies: Keep GPT-5.4 or Claude in the router for users who cannot send data to Xiaomi clouds.
- Pair with your tools: Use Healer as the perception front-end, Hunter as the planner, and close the loop with your existing RPA/agent frameworks.
FAQ
Q: Are Hunter Alpha and Healer Alpha still free?
A: For now, yes—OpenRouter keeps them at $0 while Xiaomi collects telemetry. Expect usage-based pricing once enterprise contracts roll out.
Q: Do they support custom fine-tuning?
A: Not yet. Xiaomi exposes API knobs (temperature, tool routing, system prompts) but no weight-level customization.
Q: How do they compare to GPT-5.4?
A: Xiaomi positions MiMo-V2-Pro as “Claude Opus 4.6-grade” reasoning at half the price; independent benchmarks show parity on PinchBench but not yet on math-heavy tasks.
Q: Can we self-host?
A: Xiaomi has not shipped self-hosting rights. All access flows through Xiaomi/Kingsoft clouds or OpenRouter proxy endpoints.
