Xiaomi releases two open source AI models optimized for agentic tasks

Xiaomi has launched two new open source artificial intelligence models: MiMo-V2.5 and MiMo-V2.5-Pro. Carl Franzen reports for VentureBeat that both models are available under the permissive MIT License, meaning businesses and developers can use, modify, and deploy them commercially without restrictions.

The models are designed for so-called agentic tasks, where an AI system autonomously completes multi-step work on a user’s behalf, such as managing emails, running social media accounts, or writing and publishing marketing content.

According to Xiaomi’s benchmarks, MiMo-V2.5-Pro achieves a 63.8 percent success rate on agentic tasks while using 40 to 60 percent fewer tokens than comparable models from Anthropic, Google, and OpenAI. Fewer tokens means lower costs, which matters as more services move to usage-based billing.

Xiaomi also released data on tasks the Pro model completed autonomously. It built a working compiler from scratch in 4.3 hours and developed an 8,192-line video editing application over 11.5 hours.

Pricing starts at $0.40 per million input tokens for the base model. The Pro version costs $1.00 per million input tokens for context windows up to 256,000 tokens.

Both models are available for download on Hugging Face. Xiaomi is also offering a 100 trillion free token grant to encourage developers to experiment with the models.