Hunter Alpha Unmasked: Not DeepSeek V4, but Xiaomi's MiMo-V2-Pro
TL;DR: „Hunter Alpha is Xiaomi's MiMo-V2-Pro – not DeepSeek V4. The real V4 is expected late April 2026 and will be the first frontier model running on Huawei Ascend chips."
— Till Freitag⚡ Update April 13, 2026: Hunter Alpha was confirmed on March 18 as Xiaomi's MiMo-V2-Pro – it was never DeepSeek V4. This article has been substantially revised with corrected facts on both models.
What Happened – and What We Got Wrong
On March 11, 2026, an anonymous model called Hunter Alpha appeared on OpenRouter – 1 trillion parameters, 1 million token context, completely free. The community (ourselves included) assumed: this is DeepSeek V4 in a stealth test.
We were wrong.
On March 18, Xiaomi's AI division MiMo confirmed that Hunter Alpha was an early internal test build of MiMo-V2-Pro – the company's flagship LLM. The team is led by Luo Fuli, a former core contributor at DeepSeek who joined Xiaomi in late 2025. That explains the architectural similarities that led to the confusion.
The Corrected Facts: MiMo-V2-Pro
| Property | Value |
|---|---|
| Model | Xiaomi MiMo-V2-Pro |
| Parameters (total) | >1 Trillion (1T) |
| Parameters (active) | ~42 Billion |
| Context Window | 1,048,576 Tokens (~1M) |
| Price (up to 256K tokens) | $1 / MTok Input · $3 / MTok Output |
| Price (up to 1M tokens) | $2 / MTok Input · $6 / MTok Output |
| AI Intelligence Index | ~50 (Rank 8 globally, Rank 2 China) |
| ClawEval Score | 61.5 (approaching frontier models) |
| Open Source | Planned after stabilization |
| Creator | Xiaomi / MiMo AI Team |
Important: The original free phase on OpenRouter was a limited test run. MiMo-V2-Pro is now commercially priced.
Why the Confusion Was Understandable
The evidence was real – only the conclusion was wrong:
- 1T parameters → MiMo-V2-Pro does have >1T parameters, exactly what was expected for DeepSeek V4
- MoE architecture → Both use Mixture of Experts
- Q1 2026 timing → DeepSeek V4 was expected for Q1, Hunter Alpha appeared in March
- "I am a Chinese AI model" → The model identified itself as Chinese with a May 2025 training cutoff
- Luo Fuli → The team lead came directly from DeepSeek – same architectural DNA
The lesson: in a market where multiple Chinese companies are building 1T-parameter models, "it looks like DeepSeek" is no longer proof.
So Where Is the Real DeepSeek V4?
The real DeepSeek V4 hasn't launched yet – but late April 2026 is the current target. Here's what we know as of today:
| Property | Expected Value |
|---|---|
| Parameters (total) | ~1 Trillion (1T) |
| Parameters (active) | ~32–37 Billion |
| Architecture | MoE + Engram (conditional memory) |
| Context Window | ~1M Tokens (not officially confirmed) |
| Hardware | Huawei Ascend Chips |
| SWE-bench Score | ~81% (per leaks) |
| Pricing | Expected aggressive (DeepSeek style) |
| Status | V4-Lite already in API testing |
The Huawei Story Is the Real News
Reuters confirmed on April 3: DeepSeek V4 will run on Huawei Ascend processors – not NVIDIA or AMD. DeepSeek deliberately denied Western chip makers early optimization access while giving that window to Chinese domestic manufacturers.
If V4 achieves frontier performance on Huawei chips, it's a direct counterargument to the thesis that US export controls are meaningfully slowing China's AI development.
Delayed Twice, Now Close
DeepSeek V4 was originally expected in February, then slipped to March. Since early April, Reuters reports a launch "in the next few weeks." V4-Lite is already being stress-tested on API infrastructure – a typical signal that the full launch is imminent.
What Changes in Our Assessment
What Still Holds
- The cost equation is shifting. Xiaomi's aggressive pricing and DeepSeek's expected rates confirm: AI inference is becoming a commodity.
- Agents are getting more powerful. MiMo-V2-Pro is explicitly optimized for agents (ClawEval 61.5). DeepSeek V4 is expected to push this further.
- Flexibility remains king. Hunter Alpha shows: the next relevant model can come from anywhere – Xiaomi, DeepSeek, or an entirely new player.
What's New
- China's AI ecosystem is broader than DeepSeek. Xiaomi, Alibaba (Qwen), ByteDance – a smartphone company built a frontier model. That changes the competitive dynamics.
- Talent migration matters. Luo Fuli's move from DeepSeek to Xiaomi shows: architectural know-how is portable. That makes talent, not just compute, a strategic asset.
Practical Usage: MiMo-V2-Pro Today
MiMo-V2-Pro is still available on OpenRouter – now under its real name:
Model ID: xiaomi/mimo-v2-proPerfect for OpenClaw, Cursor, Continue, or any application that supports API keys.
| Scenario | Recommendation |
|---|---|
| Prototyping & experiments | ✅ MiMo-V2-Pro |
| Code generation (non-sensitive) | ✅ MiMo-V2-Pro |
| Agentic tasks (multi-step) | ✅ MiMo-V2-Pro (ClawEval 61.5) |
| Sensitive customer data | ❌ Claude or local model |
| GDPR-relevant processing | ⚠️ Review – Xiaomi servers, not direct to China |
⚠️ Privacy Update
MiMo-V2-Pro runs on OpenRouter servers. The privacy situation has improved compared to the anonymous phase:
- Known provider: Xiaomi is a publicly listed company with clear accountability
- OpenRouter as intermediary: Data doesn't flow directly to China
- However: OpenRouter's logging policies still apply. For sensitive business data, a local model or Claude remains the better choice.
Bottom Line
Hunter Alpha was a fascinating chapter – and a lesson in humility. The speculation was understandable, but wrong. Xiaomi's MiMo-V2-Pro is an impressive model in its own right, and DeepSeek V4 is still on the horizon.
The two key takeaways:
- China's AI ecosystem is more diverse than most assumed. It's not just DeepSeek.
- DeepSeek V4 on Huawei chips will, if it launches in late April, redefine the geopolitical AI debate.
We'll keep tracking this – and correct ourselves when the facts demand it.
→ Learn more about our AI services → What is OpenClaw? → Open-Source LLM Comparison 2026
Related Articles
Deep DiveChina's AI Offensive: From Hunter Alpha to DeepSeek V4 on Huawei Chips
An anonymous 1T model, a DeepSeek mix-up, and the reveal that Xiaomi was behind it. Meanwhile, DeepSeek V4 on Huawei chi…
Read more
Deep DiveOpen-Source LLMs Compared 2026 – 25+ Models You Should Know
From Llama to Qwen to Gemma 4: all major open-source LLMs at a glance – with GitHub stars, parameters, licenses, and cle…
Read moreDeep DiveOpen-Source LLMs Compared 2026 – 25+ Models You Should Know
From Llama to Qwen to Gemma 4: Every major open-source LLM at a glance – with GitHub stars, parameters, licenses, and cl…
Read more
Gemma 4 12B Coder: Local Code Generation Becomes the Default
Google ships the Gemma 4 12B Coder — the specialized coding variant of the Gemma 4 stack. 12B parameters in GGUF format,…
Read more
Kimi K2.5: The Chinese Open-Weight Model Behind Cursor's Composer 2
Cursor's Composer 2 is secretly built on Moonshot AI's Kimi K2.5 – a 1 trillion parameter open-weight model from Beijing…
Read more
Paperclip: If OpenClaw Is the Employee, Paperclip Is the Company
Paperclip is open-source infrastructure to run an entire AI-only company – org chart, budgets, approvals, audit trail. W…
Read morePaperclip: If OpenClaw Is the Employee, Paperclip Is the Company
Paperclip is open-source infrastructure that lets you run an entire AI company — org chart, budgets, approvals, audit tr…
Read more
Ponytail: The Best Code Is the Code You Never Wrote
A dev built Ponytail because his AI agents wrote 500 lines for a 5-line problem. The result: 80-94% less code, 47-77% ch…
Read more
PewDiePie's Odysseus: The real question isn't AI sovereignty – it's the AI workplace
PewDiePie's open-source project Odysseus hit 30,000 GitHub stars in 48 hours. The more interesting question behind it: w…
Read more