New Chinese AI Model Is INSANE! (FREE & Open Source)
Xiaomi, the phone company, released MiMo V2.5 Pro, a free and open-source AI model under MIT license that outperforms Claude Opus and DeepSeek on agentic benchmarks. The model features a mixture-of-experts architecture with up to 1 trillion total parameters and a 1 million token context window. Users can access it via Hugging Face, run it locally through LM Studio, or test it free on Xiaomi's website.
Summary
The video covers the surprise release of Xiaomi's MiMo V2.5 Pro, announced on April 22nd, just days before the video was published. The presenter highlights that Xiaomi is primarily known as a phone company, making this AI release unexpected. The model is MIT-licensed, meaning it can be downloaded, fine-tuned, used commercially, and built upon at no cost.
The presenter walks through the technical architecture of the model: MiMo V2.5 Pro is a mixture-of-experts (MoE) model with 1 trillion total parameters but only 42 billion activated parameters during inference, making it more efficient to run locally than its full size would suggest. The smaller MiMo V2.5 Base variant has 310 billion total parameters with 15 billion activated. The Pro version supports a 1 million token context window, while the Base version has a smaller context length but is lighter on hardware requirements.
For access, the presenter demonstrates two main paths: downloading model weights from Hugging Face for local use, or running it through LM Studio, a free desktop application. He notes that at the time of recording, the model had just dropped and may not yet be indexed in LM Studio's model browser, but expects availability within hours. Users who don't want to run it locally can test it for free directly on Xiaomi's MiMo Chat website, where they can also switch between different model variants including voice/TTS models.
The presenter briefly demonstrates the model's coding capabilities, showing HTML games built with it and tested on LiveWeave, concluding it performs decently for coding tasks. On benchmarks, particularly the Claude Evaluation (an open agentic benchmark), MiMo V2.5 Pro is shown to outperform DeepSeek V4 Pro and Kimi K2.6, both of which the presenter claims to have personally tested. The video closes with a promotional segment for the presenter's AI Profit Builder community and training program.
Key Insights
- The presenter notes that MiMo V2.5 Pro comes from Xiaomi, a phone company with no prior AI model reputation, making the release surprising and notable in the context of competitive AI development.
- The presenter explains that MiMo V2.5 Pro uses a mixture-of-experts architecture with 1 trillion total parameters but only 42 billion activated during inference, making it more feasible to run locally than its full parameter count implies.
- The presenter states that on the Claude Evaluation agentic benchmark, MiMo V2.5 Pro outperforms both DeepSeek V4 Pro and Kimi K2.6, models he claims to have personally tested and considers strong agentic performers.
- The presenter clarifies that MiMo V2.5 Pro is specifically designed for agentic tasks — such as use with frameworks like Hermes and Open Claude — rather than being a general-purpose model.
- The presenter points out that MiMo V2.5 Pro supports a 1 million token context window, but notes this comes with higher hardware requirements compared to the Base variant, which trades context length for lighter local deployment.
Topics
Full transcript available for MurmurCast members
Sign Up to Access