Key Takeaways
- • Xiaomi MiMo-V2.5-Pro achieves top-tier agentic AI performance, scoring 1,578 on the Artificial Analysis GDPval-AA benchmark and resolving 57.2% of tasks on the SWE Bench Pro, positioning it as the leading open-weights model for autonomous agent tasks.
- • MIT-licensed open-source models deliver 40–60% token efficiency savings compared to competitors, with MiMo-V2.5-Pro priced at $1.00 per 1M input tokens and $3.00 per 1M output tokens, making cost-effective deployment feasible for enterprises.
- • Xiaomi ranks #3 globally by laboratory performance, behind only Anthropic and OpenAI, with MiMo-V2.5-Pro achieving #6 ranking globally in text arena (expert category) and #1 among open-source models as of April 26, 2026.
- • Permissive licensing enables production deployment at scale, with both MiMo-V2.5 and MiMo-V2.5-Pro available via Hugging Face under MIT License, allowing enterprises and developers to download, modify, and run models locally or on private cloud infrastructure.
- • Million-token context window and sparse MoE architecture target autonomous agent workflows, specifically powering systems for agentic "claw" tasks including marketing automation, email organization, account management, and software engineering workloads.
Summary
Xiaomi has released MiMo-V2.5 and MiMo-V2.5-Pro, open-source large language models optimized for agentic AI applications, marking a significant competitive entry into the autonomous agent market. The models, available under the MIT License on Hugging Face, represent Xiaomi's expansion beyond smartphones and electric vehicles into enterprise-grade artificial intelligence infrastructure. MiMo-V2.5-Pro delivers substantial performance improvements over its predecessor, MiMo-V2-Pro, across general agentic capabilities and complex software engineering tasks.
Performance metrics position MiMo-V2.5-Pro as a top-tier contender in autonomous agent benchmarks. The model scores 1,578 on Artificial Analysis's GDPval-AA benchmark for real-world agentic work tasks, establishing it as the leading open-weights model in this category. On the SWE Bench Pro benchmark for software engineering tasks, MiMo-V2.5-Pro resolves 57.2% of problems, competing directly with leading proprietary AI systems. In broader ranking systems, the model achieved #6 global position in the Text Arena (Expert) category as of April 26, 2026, while securing #1 status among all open-source models. This performance positions Xiaomi as #3 globally by laboratory ranking, behind only Anthropic and OpenAI, while ranking #1 among Chinese AI developers.
The models' technical architecture emphasizes efficiency and scalability for production environments. MiMo-V2.5-Pro incorporates a sparse mixture-of-experts (MoE) design with a 1-million-token context window, enabling extended processing for long-running autonomous agent workflows. Token efficiency represents a competitive advantage: Xiaomi's benchmarks demonstrate 40–60% token usage savings compared to competing models, a material consideration in usage-based billing environments where platforms like Microsoft's GitHub Copilot charge per token consumed. Pricing for MiMo-V2.5-Pro is set at $1.00 per 1M input tokens and $3.00 per 1M output tokens, positioning the model above average cost benchmarks but within competitive range for enterprise deployment.
The MIT License under which both models are released provides significant deployment flexibility, particularly for enterprises requiring production-grade agentic systems without vendor lock-in. Organizations can download models directly, execute local modifications, and operate instances on private infrastructure, eliminating dependency on API access and enabling cost optimization at scale. The release targets developers building autonomous coding agents, workflow automation systems, and interactive agents accessible via third-party messaging platforms—use cases broadly categorized as "claw" tasks involving agent-driven content creation, account management, email organization, and scheduling automation.
Performance metrics position MiMo-V2.5-Pro as a top-tier contender in autonomous agent benchmarks. The model scores 1,578 on Artificial Analysis's GDPval-AA benchmark for real-world agentic work tasks, establishing it as the leading open-weights model in this category. On the SWE Bench Pro benchmark for software engineering tasks, MiMo-V2.5-Pro resolves 57.2% of problems, competing directly with leading proprietary AI systems. In broader ranking systems, the model achieved #6 global position in the Text Arena (Expert) category as of April 26, 2026, while securing #1 status among all open-source models. This performance positions Xiaomi as #3 globally by laboratory ranking, behind only Anthropic and OpenAI, while ranking #1 among Chinese AI developers.
The models' technical architecture emphasizes efficiency and scalability for production environments. MiMo-V2.5-Pro incorporates a sparse mixture-of-experts (MoE) design with a 1-million-token context window, enabling extended processing for long-running autonomous agent workflows. Token efficiency represents a competitive advantage: Xiaomi's benchmarks demonstrate 40–60% token usage savings compared to competing models, a material consideration in usage-based billing environments where platforms like Microsoft's GitHub Copilot charge per token consumed. Pricing for MiMo-V2.5-Pro is set at $1.00 per 1M input tokens and $3.00 per 1M output tokens, positioning the model above average cost benchmarks but within competitive range for enterprise deployment.
The MIT License under which both models are released provides significant deployment flexibility, particularly for enterprises requiring production-grade agentic systems without vendor lock-in. Organizations can download models directly, execute local modifications, and operate instances on private infrastructure, eliminating dependency on API access and enabling cost optimization at scale. The release targets developers building autonomous coding agents, workflow automation systems, and interactive agents accessible via third-party messaging platforms—use cases broadly categorized as "claw" tasks involving agent-driven content creation, account management, email organization, and scheduling automation.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.