Xiaomi has announced the release and open-source availability of two new AI models, MiMo-V2.5 and MiMo-V2.5-Pro, under the MIT License. These models are designed to help developers build more cost-effective AI agents capable of handling longer tasks like coding and automation workflows. Both models support a large 1-million-token context window, allowing them to manage extensive interactions and data within a single session.

Features and Capabilities of the MiMo Models

The MiMo-V2.5-Pro is tailored for complex tasks such as coding and multi-step automation, making it suitable for enterprise-grade AI applications. In contrast, the native MiMo-V2.5 is a versatile, omnimodal model that can process not just text, but images, videos, and audio, providing a broader range of multimedia support. Xiaomi highlighted that these models use a sparse mixture-of-experts (MoE) design, which helps reduce compute costs by activating only a subset of its parameters during each request.

The MiMo-V2.5 has 310 billion parameters but only activates about 15 billion per request, while the Pro version boasts 1.02 trillion parameters, with 42 billion activated at a time. This design enables efficient processing of long contexts and complex workloads. The Pro model’s hybrid attention mechanism significantly cuts down on KV-cache storage, making it more efficient for lengthy tasks. Xiaomi reports that the models perform well in long-horizon tests, such as completing a Rust compiler in just over four hours across multiple tool calls and generating extensive video editing software after hours of autonomous work.

Open-Source Licensing and Industry Impact

By releasing these models under the MIT License, Xiaomi is allowing commercial use, further training, and fine-tuning without additional permissions. This openness is seen as a strategic move to encourage adoption among enterprise developers, especially as AI workloads grow more demanding and costly. The license’s flexibility means businesses can modify and deploy these models freely, which is a rare feature in today’s AI landscape dominated by proprietary solutions.

Industry experts believe this could make Xiaomi’s models attractive for high-volume, high-token workloads. Tulika Sheel, senior VP at Kadence International, pointed out that the MIT License’s permissiveness might drive more enterprises to experiment with Xiaomi’s offerings. She also noted that the models’ efficiency in token usage could translate into lower costs compared to other advanced AI models like Claude, Gemini, or GPT variants, which tend to consume more tokens for similar tasks.

Performance benchmarks are promising, with Xiaomi citing tests like the ClawEval, where MiMo-V2.5-Pro achieved a 64% pass rate using only about 70,000 tokens per trajectory—significantly fewer than comparable models. Such efficiency could be crucial for applications requiring extensive reasoning or multi-step processes, such as automated coding or complex decision-making workflows.

Implications for Enterprise Adoption

Whether Xiaomi’s open-source models will gain widespread use among enterprises depends on how they compare in performance, cost, and risk. Industry analysts suggest that these models might not replace the leading proprietary models entirely but could serve as cost-effective alternatives for specific high-token workloads. Lian Jye Su, from Omdia, emphasized that open models like MiMo are especially strong in high-volume, agentic tasks where the total cost of ownership becomes a key factor.

Experts also recommend that companies assess the models based on token efficiency and licensing costs. Since the models are open and can be fine-tuned, organizations may find them more flexible for customization and long-term deployment, especially when dealing with large-scale automation tasks. Pareekh Jain, CEO of Pareekh Consulting, advised enterprise teams to view MiMo-V2.5 as a tool for high-token, budget-conscious workloads rather than a direct replacement for the most advanced proprietary models.

Overall, Xiaomi’s move to open-source these models could influence how enterprises approach large-scale AI deployments in the future. The combination of high capacity, cost efficiency, and licensing freedom makes MiMo-V2.5 models a noteworthy option for organizations looking to scale their AI automation efforts without incurring prohibitive costs or licensing restrictions.

Inspired by

• https://www.infoworld.com/article/4164184/xiaomi-releases-mit%e2%80%91licensed-mimo-models-for-long%e2%80%91running-ai-agents.html

Sources

• informa.com
• pareekh.com
• gartner.com