Xiaomi MiMo-V2.5: Open-Source AI for Agentic Engineering

On April 27, 2026, the global developer landscape experienced a seismic shift as Xiaomi officially open-sourced its most ambitious AI endeavor to date: the Xiaomi MiMo-V2.5 series. Comprised of two distinct models—the native omnimodal MiMo-V2.5 and the agentic specialist MiMo-V2.5-Pro—this release marks a definitive departure from the “closed-door” culture of Western AI laboratories. By releasing these trillion-scale weights under the permissive MIT License, Xiaomi has effectively commoditized frontier-level reasoning, providing a blueprint for the next generation of autonomous, long-horizon software agents.

The “ninja” appeal of this release lies in its ruthless efficiency. While industry titans like OpenAI and Google have moved toward increasingly opaque, subscription-heavy models, Xiaomi has delivered a locally runnable alternative that matches the performance of GPT-5.4 and Claude 4.6 while consuming 40–60% fewer tokens. For developers building at the edge of autonomy, Xiaomi MiMo-V2.5 represents more than just a model; it is a declaration of independence for the open-source community, offering the technical depth required to sustain thousands of tool calls without the cognitive collapse common in smaller open-weight predecessors.

The Two-Pronged Pincer: Xiaomi MiMo-V2.5 and the Pro Variant

Xiaomi has structured this release as a “two-pronged pincer” strategy to cover the entirety of the modern AI workload spectrum. Each model is built on a Sparse Mixture-of-Experts (MoE) architecture, but they are tuned for radically different outcomes:

• MiMo-V2.5 (The Omni Specialist): This model is a native omnimodal engine with 310 billion total parameters (15 billion active). It is designed to “see, hear, and act” within a single unified architecture. Unlike older models that relied on external plug-in encoders, the V2.5 processes text, images, video, and audio natively, making it a master of multimodal perception and basic agentic tasks.
• MiMo-V2.5-Pro (The Agentic Specialist): The flagship of the series, the Pro version is a 1.02-trillion-parameter MoE model with 42 billion active parameters. It is specifically engineered for long-horizon coherence and complex software engineering. This model is the “ninja” of the bunch, trained specifically to manage the extreme “action spaces” required for autonomous coding and complex tool orchestration.

Both models support a massive 1-million-token context window, a feat made possible by Xiaomi’s proprietary architectural optimizations. This allows the Pro model to ingest entire multi-repo codebases or thousands of pages of documentation while maintaining the precision needed to execute multi-step workflows spanning thousands of individual tool calls.

Architectural Mastery: Hybrid Attention and Multi-Token Prediction

At the heart of the Xiaomi MiMo-V2.5 series is a sophisticated Hybrid Attention Architecture that solves the “KV-cache explosion” problem typical of long-context models. By interleaving Sliding Window Attention (SWA) and Global Attention (GA) at a 6:1 ratio, Xiaomi has achieved a 7x reduction in KV-cache storage requirements. This means the model can maintain “attention sinks” that anchor its focus across the 1M token span without requiring the massive hardware overhead of its competitors.

Three-Layer Multi-Token Prediction (MTP)

To address the latency issues inherent in trillion-parameter models, Xiaomi integrated three lightweight Multi-Token Prediction (MTP) modules. Standard LLMs predict one token at a time; MiMo-V2.5 predicts three tokens simultaneously during the inference phase. This triples the output speed and significantly accelerates the “rollout” phase during reinforcement learning (RL) training. For developers, this translates to an agent that doesn’t just think better, but responds with the near-instantaneous speed required for real-time collaboration.

MOPD: Multi-Teacher On-Policy Distillation

The Pro model’s superior performance in agentic tasks is largely attributed to a training regimen known as Multi-Teacher On-Policy Distillation (MOPD). During post-training, the model was refined by “learning” from multiple frontier teachers (including internal versions of MiMo-V2 and early GPT-5 clusters) across domain-specific reinforcement learning cycles. This distilled the reasoning capabilities of the world’s largest models into a more efficient, 42B active parameter footprint, enabling the Pro version to hit a GDPVal-AA Elo of 1581, effectively tying it with Claude 4.6.

Long-Horizon Coherence in Action: Real-World Benchmarks

Benchmarks such as MMLU or GSM8K are increasingly viewed as “solved” by frontier models. To prove the power of Xiaomi MiMo-V2.5-Pro, Xiaomi released data on high-complexity, real-world tasks that require sustained focus and rigorous logic over hours of autonomous operation.

1. The SysY Compiler Challenge: In a documented case study, MiMo-V2.5-Pro was tasked with building a complete SysY compiler in Rust from scratch. This involved creating a lexer, a parser, and a RISC-V assembly backend. The model completed the task in 4.3 hours, passing all 233 hidden test cases. It managed 672 tool calls without losing context or introducing regressions—a level of persistence that typically requires a senior human engineer days to achieve.
2. The Video Editor Web App: Demonstrating its omnimodal and engineering prowess, the model developed a full-featured video editor web app. The final build consisted of 8,192 lines of code, featuring a multi-track timeline, cross-fades, and an export pipeline. This required 1,868 tool calls across 11.5 hours of autonomous work, showcasing the model’s ability to “plan-do-review” in a recursive loop.

These feats are validated by its scores on SWE-bench Pro (57.2) and ClawEval (63.8), placing it at the very top of the Pareto frontier of performance versus efficiency. In the “Claw” task category—where agents must use third-party tools to schedule meetings, organize emails, and publish marketing content—the Xiaomi MiMo-V2.5 series consistently outperforms models with twice its active parameter count.

The Efficiency Advantage: Slashing Token Costs by 60%

The most disruptive element of the Xiaomi MiMo-V2.5 release for the enterprise sector is its token efficiency. In agentic workflows, the number of tokens consumed during “thought cycles” often leads to astronomical costs in closed-source ecosystems. Xiaomi’s benchmarks indicate that MiMo-V2.5-Pro reaches frontier-tier results using 40–60% fewer tokens per trajectory than GPT-5.4.

This efficiency stems from the model’s “Action Space” optimization. Because the model was trained on agent-specific trajectories, it has learned to be concise in its tool calls and reasoning chains. While a general-purpose model might “over-think” a simple file-write operation, the MiMo-V2.5-Pro executes with surgical precision. This makes it an ideal candidate for local deployment, where hardware constraints are a constant factor.

Incentivizing the Ecosystem

To ensure rapid adoption, Xiaomi announced the “One Quadrillion Token Creator Incentive Program.” Under this initiative, the company is distributing token credits worth millions of dollars to global developers. This move aims to seed the market with “MiMo-native” agents, encouraging developers to build on their stack rather than staying locked into the “buffet-style” subscription models of US-based labs that often hide their most capable models behind high-tier paywalls.

Data Sovereignty and the MIT License

In a world where data privacy is becoming the primary friction point for enterprise AI adoption, Xiaomi’s choice of the MIT License is a strategic masterstroke. By allowing commercial use, modification, and local hosting without additional authorization, Xiaomi is targeting the “regulated Western organizations” that are wary of sending proprietary data to third-party APIs.

Xiaomi MiMo-V2.5 can be deployed within a private cloud or on-premise hardware using standard inference frameworks like vLLM and SGLang. This provides “Data Sovereignty” for industries like finance, healthcare, and defense, where the security of the prompt is as valuable as the accuracy of the output. The model’s 4-bit quantization support further enables it to run on consumer-grade hardware, such as workstations equipped with the latest NVIDIA or AMD GPUs, democratizing access to trillion-scale intelligence.

Conclusion: The Dawn of the Agentic Era

The release of the Xiaomi MiMo-V2.5 and V2.5-Pro on April 27, 2026, represents the maturation of the open-source AI movement. It is no longer enough for an open-weight model to merely “chat” as well as a closed one; it must now “act” as effectively. By mastering long-horizon coherence and delivering it in a token-efficient, MIT-licensed package, Xiaomi has forced the industry to rethink the value of the proprietary API.

For the “ninja” developer, the message is clear: the tools to build truly autonomous, multi-step AI agents are now in the public domain. Whether it is constructing complex compilers in a matter of hours or managing intricate multi-modal workflows, the MiMo series proves that the gap between open and closed research is effectively closed. As we move further into 2026, the success of an AI strategy will likely be measured not by the size of the subscription budget, but by the creativity of the agents built on these powerful, open foundations.