With over a trillion parameters and top-tier coding capabilities, the new flagship model shifts the AI paradigm from talking to doing.
- Trillion-Parameter Powerhouse: Built on a highly efficient architecture, MiMo-V2-Pro features over 1T parameters (42B active) and a massive 1M-token context window, ranking 8th globally and 2nd among Chinese LLMs.
- Built for Action, Not Just Words: Designed as the core engine for agentic frameworks like OpenClaw, the model moves beyond generating text to orchestrating complex workflows, dominating serious software engineering and front-end development tasks.
- Battle-Tested and Accessible: Successfully soft-launched under the anonymous codename “Hunter Alpha” to massive developer acclaim, the MiMo-V2-Pro API is now publicly available, complete with free trial access through major developer tool partnerships.
For the past few years, the artificial intelligence industry has been captivated by the novelty of conversation. We built models that could answer questions, write poems, and generate polished demonstrations. But the future of frontier intelligence is not just about what an AI can say; it is about what it can do. Enter Xiaomi’s MiMo-V2-Pro, a flagship foundation model engineered specifically for real-world agentic workloads. By generalizing its capabilities from simple coding tasks to complex, multi-step actions, MiMo-V2-Pro is stepping up to serve as the brain behind the next generation of autonomous systems.
A Foundation Scaled for Efficiency
To build a model capable of orchestrating complex workflows, Xiaomi had to scale both model size and compute power without sacrificing speed. The result is a trillion-parameter behemoth. MiMo-V2-Pro boasts over 1T total parameters with 42B active at any given time—making it roughly three times larger than its predecessor, MiMo-V2-Flash.
Despite its massive scale, it maintains high inference efficiency by inheriting and upgrading Xiaomi’s Hybrid Attention mechanism, pushing the hybrid ratio from 5:1 up to 7:1. Paired with a lightweight Multi-Token Prediction (MTP) layer for lightning-fast generation and a staggering 1M-token context window, the model is built to handle incredibly dense information streams. The industry has already taken notice: according to the Artificial Analysis Intelligence Index—a leading global benchmark for comprehensive AI intelligence—MiMo-V2-Pro currently ranks 8th worldwide and stands as the 2nd most powerful Chinese LLM.
The “Hunter Alpha” Phenomenon
The true test of an agentic model is how it performs in the wild, outside of sterile benchmark environments. One week prior to the official announcement, Xiaomi secretly listed an early internal test build of MiMo-V2-Pro on OpenRouter, the world’s largest API aggregation platform, under the anonymous codename “Hunter Alpha.”
The response was explosive. Call volume grew steadily, sending Hunter Alpha to the top of the daily charts for multiple consecutive days and quickly surpassing 1T tokens in total usage. Interestingly, the top applications driving this volume were all coding-focused tools, validating the model’s reliability in practical development environments. Following this week of rigorous, real-world stress testing, the Xiaomi team further optimized the model, significantly improving its long-context capabilities and stability in agent scenarios before its official release.
Mastering the “Claw”: The Native Brain of Open Systems
MiMo-V2-Pro is not designed to operate in a vacuum; it is built to power the frameworks developers are already using. It serves as the native brain for OpenClaw, a general-purpose agent framework that is rapidly gaining traction in the open-source community. Because an agentic system’s ceiling is dictated by its underlying model, MiMo-V2-Pro was fine-tuned via Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) across diverse and complex agent scaffolds.
This deep optimization gives the model unparalleled tool-calling accuracy and multi-step reasoning skills. On standard open-source evaluation benchmarks like PinchBench and ClawEval, MiMo-V2-Pro achieves globally leading results. Its performance on general agent tasks approaches that of Claude Opus 4.6, and its massive 1M-token context window ensures it can comfortably ingest and process high-intensity, real-world application flows without losing the plot.
Beyond Vibe Coding: A Partner in Software Engineering
When it comes to software engineering, MiMo-V2-Pro leaves basic “vibe coding” behind. In deep evaluations conducted by Xiaomi’s internal engineering teams, the model showcased advanced code intelligence that rivals the best in the business. It actually surpasses Claude 4.6 Sonnet in dedicated coding abilities, demonstrating elegant code style, efficient problem-solving paths, and a remarkable aptitude for system design and task planning.
In frontend development scenarios, its end-to-end completion is particularly striking. Integrated with OpenClaw, MiMo-V2-Pro can generate polished, fully functional, and visually appealing web pages from a single query, perfectly balancing aesthetics with practical usability.
To ensure developers can harness this power immediately, the MiMo-V2-Pro API is now publicly available with tiered pricing based on usage. Furthermore, Xiaomi has partnered with five major agent development frameworks—OpenClaw, OpenCode, KiloCode, Blackbox, and Cline—to offer developers worldwide one week of free API access.
The Road to General Intelligence
MiMo-V2-Pro represents a massive milestone in Xiaomi’s pursuit of Artificial General Intelligence (AGI), moving the industry decisively from chat to action. But the journey doesn’t end here. The capability boundaries and system robustness of this new flagship need to be validated, stretched, and refined by developers building in complex, unknown environments.
The Xiaomi MiMo Team is committed to maintaining a rapid pace of research and engineering iteration. Going forward, their core focus will be tackling high-complexity reasoning and long-horizon task planning. By systematically improving the model’s ability to generalize and make decisions in unpredictable scenarios, MiMo-V2-Pro is laying the groundwork for a truly autonomous, intelligent future.