An Anonymous Model Takes Over OpenRouter
On March 11, 2026, a model called “Hunter Alpha” appeared on OpenRouter with no company name, no model card, and no announcement. Just an API endpoint priced so low it seemed like a mistake.
Within days, Hunter Alpha had processed over one trillion tokens. Developers and researchers worldwide discovered the same thing independently: the model was exceptionally capable, absurdly cheap, and completely unidentified. The AI community benchmarked it, probed it with prompt injections to reveal its origin, analyzed its tokenizer fingerprint, and debated theories across X, Reddit, and Hacker News.
The leading consensus pointed to DeepSeek quietly testing a next-generation system. Nobody guessed Xiaomi.
How the Community Tried to Identify Hunter Alpha
The detective work produced several technical clues. Tokenizer analysis showed a vocabulary distribution distinct from OpenAI, Anthropic, Google, or Meta models, but with structural similarities to Chinese language models. Response latency patterns suggested a Mixture of Experts (MoE) architecture with a very high total parameter count but modest active parameters per query. Users confirmed that the model maintained coherent performance up to approximately one million tokens, placing it among the most capable context windows available anywhere.
A second anonymous model, “Healer Alpha,” appeared alongside Hunter Alpha on OpenRouter. The community eventually identified it as MiMo-V2-Omni, Xiaomi’s multimodal model supporting text, image, and audio reasoning.
By the end of the anonymous period, Hunter Alpha had logged 1.27 million requests, consumed 114.6 billion prompt tokens, and generated 563.8 billion completion tokens — roughly 500 billion tokens per week at peak usage.
Xiaomi Confirms the Identity
On March 18, Xiaomi’s MiMo AI team officially announced that Hunter Alpha was an “early internal test build” of MiMo-V2-Pro. The reveal came with benchmark results verified by third parties and a pricing commitment maintaining the rates established during the anonymous phase.
The reasoning was straightforward. If MiMo-V2-Pro had launched under the Xiaomi name, evaluations would have been colored by assumptions about what a smartphone and electronics company could build. By launching anonymously, the team ensured that over one trillion tokens of real-world usage validated the model on pure merit before brand perception entered the conversation.
The strategy generated more organic attention, independent benchmarking, and media coverage than any traditional product launch could have achieved.
Advertisement
Inside the Architecture
MiMo-V2-Pro uses a Mixture of Experts architecture with approximately one trillion total parameters, of which 42 billion are active during any single forward pass. This makes it roughly three times the scale of its predecessor, MiMo-V2-Flash (309 billion total parameters, 15 billion active), while keeping inference costs manageable through sparse activation.
The model inherits a Hybrid Attention mechanism from MiMo-V2-Flash, with the ratio increased from 5:1 to 7:1, enabling the massive one-million-token context window without the quadratic compute scaling that standard transformers face. A lightweight Multi-Token Prediction (MTP) layer accelerates generation speed.
MiMo-V2-Pro is deeply optimized for agentic use cases — multi-step reasoning, sustained tool calling, long-horizon planning, and autonomous error recovery across complex scaffolds like OpenClaw.
Benchmark Performance and Pricing
On ClawEval, which measures performance in agent scaffolds including multi-turn tool use and long-horizon planning, MiMo-V2-Pro scored 61.5, placing it third globally behind Claude Opus 4.6 (66.3) and ahead of GPT-5.2 (50.0). On the Artificial Analysis Intelligence Index, it ranked eighth worldwide and second among Chinese models with a score of 49. On SWE-bench Verified, it achieved 78.0 percent, approaching Claude Opus 4.6 at 79.6 percent.
The pricing makes these numbers strategically significant. MiMo-V2-Pro costs one dollar per million input tokens and three dollars per million output tokens for context up to 256K tokens, doubling to two dollars and six dollars for the full one-million-token context. Compared to GPT-5.4 at ten dollars per million input tokens and thirty dollars per million output tokens, Xiaomi’s model runs at roughly one-fifth to one-sixth the cost for comparable frontier-class performance.
The Team and the Talent Shift
MiMo-V2-Pro was built by a team led by Luo Fuli, a former core contributor to DeepSeek’s breakthrough models. Born in 1995, Luo gained recognition at Peking University with eight papers at a single ACL conference in 2019. She joined Alibaba’s Damo Academy, then moved to High-Flyer Quant (the hedge fund behind DeepSeek), where she contributed to the development of DeepSeek-V2. Xiaomi CEO Lei Jun recruited her in late 2025 at a reported annual salary of approximately 10 million yuan (roughly 1.4 million dollars).
Luo’s move from DeepSeek to Xiaomi represents a broader trend: AI research talent in China is becoming more mobile. The expertise required to train trillion-parameter models is diffusing across more organizations, at least within China’s AI ecosystem. When individual researchers carry proven methodology for building frontier models efficiently, the number of companies capable of reaching the frontier grows rather than shrinks.
What This Signals for the AI Market
MiMo-V2-Pro challenges two assumptions that have shaped the frontier AI landscape. The first is that only American companies or purpose-built AI labs can produce globally competitive models. Xiaomi is primarily a consumer electronics and electric vehicle company, yet it delivered a system that outperforms GPT-5.2 on agentic benchmarks at a fraction of the cost.
The second assumption is that frontier AI requires frontier pricing. Chinese AI labs — first DeepSeek with its efficiency-first philosophy, now Xiaomi with hardware-software co-optimization — have consistently delivered better performance per dollar than Western counterparts. Export controls on advanced AI chips forced Chinese labs to innovate on efficiency, and what began as a constraint has become a competitive advantage.
Xiaomi has indicated it plans to open-source a variant of MiMo-V2-Pro “when the models are stable enough,” following the precedent set by MiMo-V2-Flash (released on HuggingFace in December 2025). Even without open weights, the low API pricing already makes frontier capabilities accessible to developers and organizations that were priced out of GPT-5.x or Claude Opus 4.6 tiers.
Frequently Asked Questions
How did Xiaomi build a frontier AI model when the company is known for smartphones?
Xiaomi recruited Luo Fuli and other researchers from DeepSeek, the Chinese lab that demonstrated how to build frontier models with exceptional efficiency. Luo, who contributed to DeepSeek-V2, brought proven methodology for training large-scale models at lower compute cost. Combined with Xiaomi’s revenue base and infrastructure investment, the team scaled the MiMo architecture from MiMo-V2-Flash (309 billion parameters) to MiMo-V2-Pro (one trillion parameters) while maintaining the efficiency-first design philosophy that keeps inference costs low.
Is MiMo-V2-Pro really five times cheaper than competitors?
Yes, and the pricing appears sustainable. At one dollar per million input tokens and three dollars per million output tokens (up to 256K context), MiMo-V2-Pro costs roughly one-fifth what GPT-5.4 charges. The cost advantage comes from the MoE architecture (only 42 billion of one trillion parameters activate per query) and the 7:1 hybrid attention mechanism that reduces compute requirements for long-context inference. The tradeoff is a less mature ecosystem — documentation, SDKs, and enterprise support are still catching up to OpenAI and Anthropic.
Will MiMo-V2-Pro be available as open-source weights?
Xiaomi has indicated it plans to release an open-source variant “when the models are stable enough.” The company followed this pattern with MiMo-V2-Flash, which was open-sourced on HuggingFace in December 2025. Currently, MiMo-V2-Pro is accessible only through Xiaomi’s API platform and OpenRouter. Even without open weights, the low API pricing partially addresses the accessibility gap that open-sourcing would solve.
Sources & Further Reading
- Xiaomi Stuns With New MiMo-V2-Pro LLM Nearing GPT-5.2, Opus 4.6 Performance at a Fraction of the Cost — VentureBeat
- Who’s That AI? The Mystery Model Everyone Blamed on DeepSeek Turned Out to Be Xiaomi — Technology.org
- Chinese AI Prodigy Luo Fuli Joins Xiaomi as Industry Competition for Talent Heats Up — South China Morning Post
- Xiaomi Launches Three MiMo AI Models to Power Agents, Robots, and Voice — The Decoder
- MiMo-V2-Pro: Intelligence, Performance and Price Analysis — Artificial Analysis
- MiMo-V2-Pro API Pricing and Providers — OpenRouter
