๐ญ๐ฐSCMP TechnologyโขStalecollected in 30m
Xiaomi launches top-ranked open-source agentic AI model
๐กXiaomi's new open-source model just topped agentic benchmarks, signaling a major shift in hardware-integrated AI.
โก 30-Second TL;DR
What Changed
MiMo-V2.5-Pro is optimized for agentic workflows in hardware environments.
Why It Matters
This move signals Xiaomi's intent to compete directly with global tech giants by building a proprietary, agent-focused AI stack for its hardware ecosystem. It highlights the growing trend of hardware manufacturers prioritizing on-device agentic intelligence.
What To Do Next
Evaluate the MiMo-V2.5-Pro benchmarks on Artificial Analysis to see if its agentic performance suits your specific automation use cases.
Who should care:Developers & AI Engineers
Key Points
- โขMiMo-V2.5-Pro is optimized for agentic workflows in hardware environments.
- โขRanked as the world's top open-source model for agentic tasks by Artificial Analysis.
- โขPart of Xiaomi's strategic shift to integrate AI across its smartphone and EV ecosystem.
๐ง Deep Insight
Web-grounded analysis with 20 cited sources.
๐ Enhanced Key Takeaways
- โขMiMo-V2.5-Pro is a Mixture-of-Experts (MoE) model featuring 1.02 trillion total parameters and 42 billion active parameters, designed for demanding agentic, complex software engineering, and long-horizon tasks.
- โขThe model incorporates a hybrid attention architecture, interleaving Sliding Window Attention and Global Attention at a 6:1 ratio, and utilizes 3-layers Multi-Token Prediction (MTP) to enhance efficiency and inference speed.
- โขTrained on an extensive dataset of 27 trillion tokens using FP8 mixed precision, MiMo-V2.5-Pro supports an impressive context length of up to 1 million tokens.
- โขMiMo-V2.5-Pro demonstrates superior token efficiency, consuming approximately 40-60% fewer tokens than comparable frontier models like Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 for equivalent capabilities on agentic benchmarks such as ClawEval.
- โขXiaomi has committed to investing over 60 billion yuan (approximately US$8.8 billion) in artificial intelligence over the next three years, underscoring its strategic focus on developing advanced AI models and integrating them across its product ecosystem.
๐ Competitor Analysisโธ Show
| Feature/Model | Xiaomi MiMo-V2.5-Pro (Open-Source) | Kimi K2.6 (Open-Source) | DeepSeek V4 (Open-Source) | Qwen 3.6 Plus (Open-Source) | GLM 5.1 (Open-Source) | Claude Opus 4.6 (Closed-Source) | GPT-5.4 (Closed-Source) |
|---|---|---|---|---|---|---|---|
| Total Parameters | 1.02 Trillion (42B active) | N/A | N/A | N/A | N/A | N/A | N/A |
| Context Window | 1 Million tokens | N/A | 128K tokens | 1 Million tokens | N/A | N/A | N/A |
| Agentic Ranking (Artificial Analysis) | Top open-source model (tied with Kimi K2.6) | Top open-source model (tied with MiMo-V2.5-Pro) | Trailing only DeepSeek and Moonshot AI among open-source in overall intelligence and coding tests | Outperforms GPT-4 on several agentic tool-use benchmarks | Rivals GPT-5.4 on structured coding tasks | 67.6% (Agentic Index) | 69.4% (Agentic Index) |
| Token Efficiency | 40-60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, GPT-5.4 for comparable capability; 42% fewer than Kimi K2.6 on ClawEval | N/A | N/A | N/A | N/A | N/A | N/A |
| License | MIT License | N/A | N/A | N/A | MIT License | Proprietary | Proprietary |
| Key Strengths | Agentic, complex software engineering, long-horizon tasks, high token efficiency, multimodal (V2.5 variant) | N/A | Frontier reasoning, long-context, tool-use efficiency | Agentic coding, reliable tool use, long context, error correction | Code generation quality, tool use | Strong agentic performance, multi-step planning | Leading agentic performance, multi-step planning |
๐ ๏ธ Technical Deep Dive
- Model Type: Mixture-of-Experts (MoE) language model.
- Parameters: 1.02 trillion total parameters with 42 billion active parameters per token.
- Architecture: Inherits hybrid attention and Multi-Token Prediction (MTP) design from MiMo-V2-Flash.
- Hybrid Attention: Interleaves Local Sliding Window Attention (SWA) and Global Attention (GA) at a 6:1 ratio with a 128-token window. This design reduces KV-cache storage by nearly 7x while maintaining long-context performance via a learnable attention-sink bias.
- Multi-Token Prediction (MTP): Equipped with three lightweight MTP modules using dense FFNs, which triples output throughput during inference and accelerates Reinforcement Learning (RL) rollouts.
- Pre-training: Trained on 27 trillion tokens using FP8 mixed precision at a native 32k sequence length.
- Context Window: Supports up to 1 million tokens.
- Post-training for Agentic Capabilities: Utilizes Supervised Fine-Tuning (SFT), large-scale agentic Reinforcement Learning, and Multi-Teacher On-Policy Distillation (MOPD).
- Modalities: MiMo-V2.5-Pro is text-focused for agentic reasoning, while its variant MiMo-V2.5 supports native full-modal agent capabilities covering images, audio, and video.
- Benchmark Performance: Achieves 64% Passยณ on ClawEval using approximately 70,000 tokens per trajectory, demonstrating significant token efficiency.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
Xiaomi's deep integration of MiMo-V2.5-Pro into its HyperOS will significantly enhance user experience across its "Human ร Car ร Home" ecosystem.
The model's agentic capabilities and optimization for hardware environments suggest a seamless, intelligent interaction across smartphones, EVs, and smart home devices, moving beyond traditional voice assistants to full-ecosystem AI agents.
Xiaomi's open-source strategy for MiMo-V2.5-Pro will accelerate its adoption and foster a broader developer ecosystem.
By open-sourcing a frontier-tier model with competitive performance and token efficiency, Xiaomi can attract global developer talent and potentially establish its model as a standard for agentic AI development.
Xiaomi's substantial AI investment and focus on internal manufacturing efficiency with AI will yield a significant competitive advantage in global tech markets.
The company's commitment of over $8.8 billion to AI and its deployment of AI and robotics for production efficiency and innovation speed aim to create a structural advantage in cost and speed, especially in the US-China tech competition.
โณ Timeline
2010-04
Xiaomi founded, began developing MIUI OS.
2023-10
Xiaomi upgraded its strategy to "Human ร Car ร Home" smart ecosystem.
2025-04
Xiaomi announced its venture into large-scale AI models with the introduction of Xiaomi MiMo, starting with MiMo-7B.
2025-12
MiMo-V2-Flash, an open-sourced Mixture-of-Experts model, was launched.
2026-03
Xiaomi CEO Lei Jun announced plans to invest at least US$8.7 billion in AI over the next three years. MiMo-V2-Pro and MiMo-V2-Omni were publicly introduced.
2026-04-22
Xiaomi released MiMo-V2.5-Pro and MiMo-V2.5.
๐ Sources (20)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: SCMP Technology โ