Xiaomi drops MiMo-V2-Pro: The mystery model that topped OpenRouter

Over the past week, developers using the API aggregator OpenRouter noticed an anonymous model codenamed "Hunter Alpha" quietly taking over the daily charts. It racked up over 1 trillion tokens in usage, with coders everywhere wondering who was behind it. Now, Xiaomi has come clean: it was the test build for its new flagship model, MiMo-V2-Pro.

The mystery model revealed

Officially launched on March 18, MiMo-V2-Pro is Xiaomi's dedicated foundation model engineered specifically for real-world agent workflows. The company says continuous tuning over the test week sharpened both its long-context handling and stability.

Under the hood, the model packs over 1 trillion total parameters with 42B active, roughly three times the size of its predecessor, MiMo-V2-Flash. Xiaomi kept its Hybrid Attention mechanism but bumped the ratio from 5:1 to 7:1 to balance massive scale with low latency, aided by a lightweight Multi-Token Prediction layer. It also ships with a massive 1M-token context window.

On the Artificial Analysis Intelligence Index, an industry benchmark for comprehensive model intelligence, MiMo-V2-Pro currently ranks 8th globally and 2nd among Chinese large language models.

Moving from vibe coding to real agents

Xiaomi is skipping generic chatbot chatter and focusing strictly on task execution. The team pitched it as expanding frontier intelligence from "coding to claw," referencing OpenClaw, the booming open-source autonomous agent framework.

MiMo-V2-Pro was fine-tuned with SFT and reinforcement learning across complex agent scaffolds. On OpenClaw's benchmark suites, PinchBench and ClawEval, Xiaomi says the model posted top-tier scores. In terms of raw developer capability, Xiaomi claims its coding ability surpasses Claude 4.6 Sonnet, while its general agentic performance approaches Claude Opus 4.6.

During the anonymous Hunter Alpha rollout, the top-consuming apps were all coding assistants, with community feedback frequently favoring it over Sonnet. To accelerate adoption, Xiaomi is partnering with five popular agent frameworks—OpenClaw, OpenCode, KiloCode, Blackbox, and Cline—to give developers worldwide one week of free API access.

Why it matters

We think Xiaomi is playing its AI cards remarkably well here.

Hardware makers usually treat AI as an internal feature to boost phone or gadget sales, rarely venturing into the deep end of developer tooling. By building a 1T-parameter model optimized for agentic execution—and proving its chops anonymously on OpenRouter before taking credit—Xiaomi proved it can compete on raw developer utility rather than corporate hype.

For developers, having another heavyweight contender in the agent space is huge. Right now, the autonomous coding ecosystem is heavily reliant on Anthropic's Claude lineup. A viable alternative that supports a 1M-token context window, offers free temporary prompt cache writes, and integrates natively into frameworks like OpenClaw gives builders serious leverage.

That said, claiming parity with Claude in marketing benchmarks is different from surviving months of messy, real-world code refactoring and chaotic multi-step agent loops. But Xiaomi's aggressive engineering and ecosystem-first launch make MiMo-V2-Pro one of the most interesting model drops of the year.