Xiaomi has introduced two new open-source large language models, Xiaomi MiMo V2.5 and Xiaomi MiMo V2.5 Pro. Both models are released under the MIT License, allowing developers and enterprises to use, modify, fine-tune, and deploy them commercially.

The models are available through Hugging Face and can be deployed locally, on private cloud systems, or through Xiaomi’s API services. The company is positioning the new lineup for agentic workloads such as coding, content creation, workflow automation, scheduling, and long multi-step tasks.

According to Xiaomi’s benchmark data, MiMo V2.5 Pro achieved a 63.8% success rate on ClawEval while using around 70,000 tokens per task trajectory.

Xiaomi said this is roughly 40% to 60% fewer tokens than comparable results from models such as Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4. Lower token use could reduce operating costs for businesses running long agentic workflows.

The standard MiMo V2.5 model is designed for multimodal use cases involving text, image, and audio reasoning.

The MiMo V2.5 Pro version is focused on advanced coding, tool use, long horizon planning, and sustained multi-step execution. Xiaomi said the Pro model scored 1581 on the GDPVal-AA Elo benchmark, placing it ahead of Kimi K2.6 and GLM 5.1.

Xiaomi also shared internal autonomous task results for MiMo V2.5 Pro. The model reportedly built a full SysY compiler in Rust in 4.3 hours using 672 tool calls and passed all 233 hidden tests.

It also created an 8,192-line desktop video editor in 11.5 hours using 1,868 tool calls and improved an analog circuit design through repeated simulation loops.

MiMo V2.5 uses a Mixture-of-Experts architecture with 310 billion total parameters and 15 billion active parameters during inference.

The Pro model expands this to 1.02 trillion total parameters with 42 billion active parameters.

Both models support context windows of up to 1 million tokens, allowing them to process very large documents and long sessions.

For overseas developers, MiMo V2.5 Pro starts at $1 per million input tokens and $3 per million output tokens for context windows up to 256K. The standard MiMo V2.5 starts at $0.40 per million input tokens and $2 per million output tokens.

Xiaomi has also introduced annual Token Plan subscriptions ranging from $63.36 to $1,056, depending on usage levels. The plans include discounted API pricing and support for coding tools such as Cursor, Zed, and Claude Code.

The MIT License gives companies broad flexibility. Businesses can use the models commercially, fine-tune them with private data, and release derivative versions without common licensing restrictions.

This could make MiMo more attractive than some open-weight models that still impose commercial or usage limits.

Xiaomi added that the next generation of MiMo models is already in training, with a focus on deeper reasoning and stronger real-world grounding.

📢 For the latest Tech & Telecom news, videos and analysis join ProPakistani's WhatsApp Group now!

Follow ProPakistani on Google News & scroll through your favourite content faster!

Shares