Hunter Alpha Unmasked: Xiaomi's MiMo-V2-Pro
The mystery model that topped OpenRouter charts and sparked DeepSeek V4 rumors was actually Xiaomi's flagship AI — built by former DeepSeek researchers and positioned as the autonomous brain for smartphones and electric vehicles.
What Happened
On March 11, 2026, a powerful AI model appeared on OpenRouter with zero developer attribution. Named "Hunter Alpha," it advertised 1 trillion parameters and a 1 million token context window. The AI community immediately speculated it was DeepSeek V4 testing in stealth mode.
Eight days later, Xiaomi revealed the truth: Hunter Alpha was MiMo-V2-Pro, the smartphone giant's flagship foundation model for agentic workloads.
The Specs
| Spec | MiMo-V2-Pro | Hunter Alpha (Original) |
|---|---|---|
| Total Parameters | ~1 trillion | 1 trillion (advertised) |
| Active Parameters | 42B per token | Not disclosed |
| Architecture | MoE, Hybrid Attention (7:1) | Not disclosed |
| Context Window | 1M tokens | 1M tokens |
| Max Output | 131K tokens | 32K tokens |
| Modality | Text + Image → Text | Text + Image → Text |
| License | Proprietary (API-only) | Not disclosed |
Benchmarks
MiMo-V2-Pro ranks 8th worldwide and 2nd among Chinese LLMs on the Artificial Analysis Intelligence Index.
| Benchmark | MiMo-V2-Pro | Comparison |
|---|---|---|
| Intelligence Index | 49 | GLM-5: 50, GPT-5.2 Codex: ~49 |
| GDPval-AA (Agentic) | 1426 ELO | GLM-5: 1406, Claude Sonnet 4.6: 1633 |
| ClawEval (Agent Scaffold) | 61.5 | Claude Opus 4.6: 66.3, GPT-5.2: 50.0 |
| Hallucination Rate | 30% | MiMo-V2-Flash: 48% |
| Token Efficiency | 77M tokens | GLM-5: 109M, Kimi K2.5: 89M |
The Pro model is the highest-scoring Chinese-origin model on GDPval-AA, beating GLM-5 (1406) and Kimi K2.5 (1283). On ClawEval, it scores 61.5 — approaching Claude Opus 4.6's 66.3.
Pricing
| Context Range | Input | Output | Cache Read |
|---|---|---|---|
| Up to 256K | $1/M tokens | $3/M tokens | $0.20/M |
| 256K to 1M | $2/M tokens | $6/M tokens | $0.40/M |
Compare to Claude Opus 4.6: $15/M input, $75/M output. MiMo-V2-Pro delivers ~92% of GLM-5's intelligence at roughly 1/5th the cost of frontier Western models.
Technical Innovations
Rollout Routing Replay (R3)
MoE models suffer from routing drift — experts activated during inference differ from those used during training. R3 enforces deterministic constraints where rollout experts are strictly reused during backpropagation, eliminating performance degradation.
Multi-Token Prediction (MTP)
MiMo-V2-Flash generates at 141.9 tokens/second (nearly 2.5x the average for similar-size open-weight models). MTP predicts multiple future tokens in a single forward pass.
Multi-Teacher On-Policy Distillation (MOPD)
Domain-specialized teacher models provide dense, token-level rewards during training, improving efficiency for the Flash variant.
The Xiaomi Pivot
MiMo-V2-Pro isn't a demo project. Xiaomi positions it as the autonomous brain for:
- Smartphones (integrated OS features)
- SU7 electric vehicle (autonomous driving systems)
- Smart home ecosystem (IoT orchestration)
The MiMo team is led by Luo Fuli, a former DeepSeek researcher. Xiaomi gave Chinese chipmakers early optimization access while denying that window to NVIDIA and AMD — a strategic signal about China's AI hardware direction.
Community Reactions
"For creative writing use cases, MiMo-V2-Pro just put itself on a very short list, and in our opinion is by far the best and richest model available, beating Claude 4.6 Opus easily." — Decrypt Review
"The combination that stood out was Hunter Alpha's 1 million token context paired with reasoning capability and free access. Most frontier models with that context window come with real cost at scale." — Nabil Haouam, AI engineer
"My analysis suggests Hunter Alpha is likely not DeepSeek V4." — Umur Ozkul, independent benchmark tester (who was right)
Why This Matters
- Xiaomi isn't a traditional AI company — yet they built a frontier-class model that fooled experts.
- Agentic focus — positioned for real-world tool workflows, not chat demos.
- Cost disruption — 1/5th the cost of Claude/GPT at near-frontier performance.
- Hardware independence — optimized for Chinese chips, not NVIDIA.
- The mystery strategy worked — stealth testing generated massive attention before official reveal.
Access
- OpenRouter:
openrouter/xiaomi/mimo-v2-pro(1M context) - Xiaomi AI Studio: Official API platform
- MiMo-V2-Flash: Open weights on HuggingFace (MIT license, 309B params, 15B active)
Sources: Reuters, VentureBeat, Decrypt, Artificial Analysis, Xiaomi official launch page, OpenRouter model cards