AI Dev Weekly Extra — a special edition for breaking news that can’t wait until Thursday.

A week ago, a model called “Hunter Alpha” quietly appeared on OpenRouter. No developer attribution. No company name. Just a free, anonymous model that was suspiciously good.

Within days, the AI community had one theory: this was DeepSeek V4, stealth-testing before launch.

They were wrong. It was Xiaomi.

What happened

On March 11, Hunter Alpha showed up on OpenRouter — the API gateway that lets developers access dozens of AI models through a single endpoint. OpenRouter labeled it a “stealth model.” No one knew who built it.

The specs were wild: a 1-million-token context window, strong reasoning capabilities, and it was completely free to use. Developers started testing it, and the results were impressive enough that speculation exploded. The leading theory was that DeepSeek — the Chinese AI lab that shook global markets with its earlier models — was quietly testing V4 before an official launch.

On March 18, Xiaomi’s AI division MiMo confirmed that Hunter Alpha was actually an early internal test build of their flagship model, MiMo-V2-Pro.

The specs

MiMo-V2-Pro is a mixture-of-experts model:

• 1 trillion total parameters (but only 42 billion active during inference — the MoE trick that keeps costs down)
• 1 million token context window — matching the longest context windows available
• Hybrid attention architecture optimized for multi-step reasoning and tool use
• Led by Luo Fuli, a former DeepSeek researcher — which explains why people assumed the DeepSeek connection

The model is specifically designed for “agent” workloads — complex multi-step tasks where the AI needs to plan, use tools, and execute sequences of actions. Think: AI coding assistants, automated workflows, research agents.

Why developers should care

Three things stand out:

1. The phone company is serious about AI. Xiaomi isn’t just slapping AI features into their phones. They’re building frontier-scale models that compete with dedicated AI labs. Lei Jun (Xiaomi’s CEO) announced an additional 16 billion yuan investment in AI alongside this launch. When a company with Xiaomi’s distribution and hardware ecosystem goes all-in on AI, the developer tooling implications are massive.

2. MoE is winning. MiMo-V2-Pro joins DeepSeek, Mixtral, and others in proving that mixture-of-experts architectures deliver the best cost-to-performance ratio. 1 trillion parameters but only 42 billion active means you get near-frontier performance at a fraction of the inference cost. For developers building on top of these models, this means cheaper API calls for complex tasks.

3. The agent era is here. Xiaomi explicitly positioned MiMo-V2-Pro as an “agent model” — not a chatbot, not a code completer, but a model designed to autonomously execute multi-step workflows. This aligns with what we’re seeing from Anthropic (Claude’s computer use), OpenAI (Operator), and Google (Project Mariner). The models are moving from “answer my question” to “do this task for me.”

The DeepSeek connection

The fact that everyone assumed Hunter Alpha was DeepSeek V4 tells you something about DeepSeek’s reputation. After DeepSeek-V3 and R1 disrupted the market with their efficiency claims, the AI community is watching for V4 like hawks. Any anonymous model that performs well gets the DeepSeek label by default.

The Luo Fuli connection adds another layer. Having a former DeepSeek researcher lead Xiaomi’s AI team means there’s genuine knowledge transfer happening. The techniques that made DeepSeek efficient are now being applied at Xiaomi’s scale.

What this means for your stack

If you’re building AI-powered features, MiMo-V2-Pro is worth watching. The 1M context window and agent-focused design make it a potential alternative to Claude or GPT for:

• Long document processing
• Multi-step code generation
• Automated testing and debugging workflows
• Research and analysis pipelines

The model is available on OpenRouter now. Pricing hasn’t been finalized, but during the stealth test it was free — and Xiaomi has signaled they’ll compete aggressively on price.

I’ll be testing it against Claude and GPT for coding tasks this week. Stay tuned for the results in next Thursday’s regular edition.

---

AI Dev Weekly publishes every Thursday. Read the full series for weekly coverage of AI developer news.

Want to go deeper? Read What Is MiMo-V2-Pro? for the full explainer, or MiMo-V2-Pro vs Claude vs GPT for the head-to-head comparison.