What Happened
On March 11, 2026, an anonymous model called Hunter Alpha appeared on OpenRouter. No description, no organization name, no benchmarks. Within days it topped the platform's daily usage charts and processed over 1 trillion tokens in a week. Developers noticed it was exceptionally capable at agentic tasks but could not identify who made it.
On March 18, 2026, Xiaomi revealed that Hunter Alpha was an early test build of MiMo-V2-Pro — their flagship foundation model, built for agent workflows, with over 1 trillion total parameters and a 1 million token context window.
You probably know Xiaomi for smartphones and, more recently, electric vehicles. The MiMo-V2-Pro announcement signals something different: Xiaomi is competing directly with Anthropic, OpenAI, and Google at the foundation model level.
The Benchmark Numbers
| Benchmark | MiMo-V2-Pro | Claude Sonnet 4.6 | Claude Opus 4.6 |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 49 (top tier) | ~45 | ~62 |
| PinchBench (agent tasks) | Top tier globally | — | Near parity |
| ClawBench (coding agents) | Top tier globally | Below MiMo | Below MiMo |
| GPQA Diamond | 87% | ~80% | ~88% |
| Context window | 1M tokens | 200K tokens | 200K tokens |
| Price (input per 1M tokens) | $0.10 | $3.00 | $15.00 |
The headline claim from Xiaomi: MiMo-V2-Pro beats Claude Sonnet 4.6 on coding benchmarks and approaches Claude Opus 4.6 on agent benchmarks — at 67% lower cost than Sonnet and 99% lower cost than Opus.
On the Artificial Analysis Intelligence Index, it scores 49 against a median of 14 for models in the same price tier (under $0.15/1M tokens). It currently ranks #8 globally on that index and #1 in its price tier.
Source: Artificial Analysis and OpenRouter, verified March 2026. Note: benchmarks are preliminary — the model was released 5 days ago and independent evaluations are still being finalized.
The Architecture
Xiaomi has not published a full technical report. What is confirmed:
- Over 1 trillion total parameters — parameter count per active token not disclosed
- Hybrid ratio 7:1 — MiMo-V2-Pro uses a 7:1 hybrid attention-to-other ratio to manage the 1M token context window (up from 5:1 in the earlier MiMo-V2-Flash)
- Reasoning model — uses chain-of-thought / extended thinking before answering, similar to DeepSeek-R1 and o3
- Text-only — no image input as of launch
- SFT + RL training on agent scaffolds — explicitly fine-tuned on complex, diverse agent workflows, not just instruction following
The "Hunter Alpha" week on OpenRouter was a production test, not a demo. Xiaomi used real developer traffic to measure long-context stability and agent performance under load before the official announcement.
What It Is Built For
MiMo-V2-Pro is not optimized for chat. Xiaomi's framing is explicit: this model is designed to be the orchestration layer of multi-agent systems.
The relevant capabilities for agent builders:
Tool calling reliability — trained across diverse agent scaffolds, strong on multi-step tool invocations. Compatible with OpenClaw, Xiaomi's open-source agent framework, and standard frameworks like LangGraph and AutoGen.
Long-context coherence — the 7:1 hybrid architecture is specifically designed to maintain reasoning quality across long agent histories. A 1M context window is only useful if quality does not degrade past 200K.
Code generation — Xiaomi's demo showed MiMo-V2-Pro generating a complete, functional frontend application from a detailed spec in a single pass. The spec included: multi-column magazine grid, sepia image filters, page-turn transitions, and a monospace/serif font combination. The output matched all specifications without iteration.
Pricing and Access
| Access method | Cost | Notes |
|---|---|---|
| OpenRouter (first week) | Free | Week ended ~March 25, 2026 |
| OpenRouter (after free period) | ~$0.10/1M input | Competitive with cheapest open models |
| Xiaomi Direct API | TBD | mimo.xiaomi.com — registration required |
| Self-hosted | Not available | Weights not released |
The model is proprietary — weights are not open. This distinguishes it from Llama 3.3 and Nemotron 3 Super, which can be self-hosted. For organizations with data privacy requirements, using MiMo-V2-Pro means sending data to Xiaomi's infrastructure.
Why the Anonymous Launch Matters
Xiaomi's approach of deploying Hunter Alpha anonymously before revealing it was MiMo-V2-Pro is unusual and worth understanding.
The standard model launch process: announce, publish technical report, open API access, collect benchmark results. Xiaomi reversed it: ship to production, collect real usage data at scale (1T tokens, 500B tokens/week), then announce with production performance data rather than synthetic benchmarks.
This approach has two advantages. First, the 1T token production run is a form of benchmark that synthetic evals cannot replicate — real developer queries across real use cases. Second, the usage charts (top of OpenRouter for multiple days before anyone knew who made it) are a marketing signal that is hard to manufacture.
The playbook is reminiscent of how DeepSeek-R1 launched: strong technical claims, verified independently, underpriced relative to Western alternatives.
What It Does Not Do Well
MiMo-V2-Pro does not support image input. For multimodal agent workflows — browser automation, screenshot analysis, document parsing — you still need GPT-5.4, Claude, or Gemini.
The 32K maximum output token limit is a practical constraint for tasks that require generating long documents or large codebases in a single call.
Xiaomi's proprietary infrastructure is an unknown variable for enterprise adoption. There is no SOC 2 report, no GDPR Data Processing Agreement published, and no regional data residency option announced as of March 2026.
FAQ
Is MiMo-V2-Pro still free?
The one-week free period announced at launch ends approximately March 25, 2026. After that, access is through OpenRouter at standard pricing (~$0.10/1M input tokens) or the Xiaomi direct API. Check openrouter.ai/xiaomi/mimo-v2-pro for current pricing.
Can I use MiMo-V2-Pro with LangChain or LangGraph?
Yes. Via OpenRouter's OpenAI-compatible endpoint, use ChatOpenAI with base_url=\"https://openrouter.ai/api/v1\" and model=\"xiaomi/mimo-v2-pro\". The model supports tool calling through the standard OpenAI function calling interface.
Is this the same as MiMo-V2-Flash?
No. MiMo-V2-Flash is a smaller, faster variant. MiMo-V2-Pro is the flagship — more parameters, 7:1 hybrid ratio (vs 5:1 in Flash), better agent benchmark scores, higher cost. Think of Flash as the equivalent of Claude Haiku and Pro as the equivalent of Claude Sonnet.
Should I trust benchmarks from a 5-day-old model release?
With caution. Xiaomi's own benchmarks have not yet been independently replicated at scale. Artificial Analysis scores are based on standardized test runs, but community validation is still ongoing. The production usage data (1T tokens on OpenRouter) is the most credible signal — real developers chose to use it at scale when they did not even know who made it.
Why is a phone company building frontier AI models?
Xiaomi's strategy is vertical integration across its ecosystem: phones, EVs, smart home devices, and now the AI layer that coordinates them. A proprietary foundation model gives Xiaomi independence from OpenAI and Anthropic API pricing, and control over the AI capabilities embedded in its products. This is the same rationale behind Apple Intelligence and Google's Gemini Nano.
Sources
Next step: If you are building an agent that requires long context and strong tool-calling at low cost, test MiMo-V2-Pro via OpenRouter today using the OpenAI-compatible endpoint. Run your standard eval suite and compare against your current model on both quality and cost per task — the pricing gap is large enough that a 10-15% quality trade-off may still make economic sense at scale.