Inside Xiaomi’s Hunter Alpha: 1‑Trillion‑Parameter LLM with 1M Context and Top Global Rankings

MiMo‑V2‑Pro, announced by Xiaomi as the mysterious top‑ranked model on OpenRouter under the codename Hunter Alpha, is a large language model (LLM) with a total parameter count exceeding 1 trillion and an activation parameter count of 42 billion. Its context window reaches 1 million tokens, making it suitable for real‑world tasks that require very long inputs.

The model’s architecture inherits a mixed‑attention mechanism and increases the mixing ratio from 5:1 to 7:1, balancing ultra‑large scale with high inference efficiency. It also incorporates a lightweight MTP (Multi‑Token Prediction) layer, allowing fast generation even when processing the full 1 M‑token context.

To train and serve such a massive model, Xiaomi’s AI team co‑developed the ARL‑Tangram unified resource‑management system with Peking University. ARL‑Tangram uses a unified action‑level formula and elastic scheduling algorithm, improving average action completion time by 4.3 points and reducing training‑step duration by up to 1.5×, cutting overall compute cost and external resource consumption by 71.2%.

According to the latest Artificial Analysis evaluation, MiMo‑V2‑Pro ranks second among domestic models and eighth globally. On the PinchBench suite it scores 84.0, and on ClawEval it achieves 61.5, surpassing Gemini 3 Pro and approaching Claude Opus 4.6. In the SWE‑bench Verified test for code generation it obtains a score of 78, demonstrating strong system‑design and complex‑task planning abilities.

During blind testing on OpenRouter, the model’s total token usage quickly topped the leaderboard, with OpenClaw consuming 675 B tokens, Kilo Code 50.1 B tokens, and Roo Code 38.6 B tokens, confirming its reliability in production‑grade code‑heavy workloads.

One concrete demonstration shows the model generating a complete Three.js 3D tower‑defense game with dynamic background and upgrade path from a single natural‑language prompt, illustrating its ability to translate high‑level instructions into functional code.

The companion model MiMo‑V2‑Omni is a multimodal LLM that fuses image, video, and audio encoders into a single backbone, enabling it to perceive and act on visual, auditory, and textual inputs simultaneously. In tests it acted as an autonomous driving “brain” to anticipate hazards and performed cross‑platform price‑comparison shopping within a browser, handling tool calls and UI interactions seamlessly.

MiMo‑V2‑TTS is a speech model designed to give agents emotional expression. By providing natural‑language prompts, users can precisely control timbre, tone, and emotional nuances; the model can insert coughs, sighs, or laughter automatically based on context.

Before the official reveal, the identity of Hunter Alpha sparked widespread speculation. Some users guessed it belonged to OpenAI or DeepSeek, while others noted its strong Chinese training data and suggested a domestic origin. The community’s curiosity was highlighted when Peter Steinberger (the “father of the lobster” UI framework) asked for details on X.

On March 11 the models appeared on OpenRouter, quickly amassing over 1 trillion tokens of usage and maintaining the top spot for several days. The official announcement coincides with Xiaomi’s upcoming product launch, suggesting the models may underpin new services.

For readers interested in experimenting, the models are accessible via the following URLs:

https://mimo.xiaomi.com/mimo-v2-pro<br/>https://mimo.xiaomi.com/mimo-v2-omni<br/>https://mimo.xiaomi.com/mimo-v2-tts