Xiaomi just dropped a 1 trillion parameter language model on Hugging Face, and it's MIT licensed. MiMo-V2.5-Pro ships with a 1 million-token context window and benchmarks that compete with top proprietary models. It scores 78.9 on SWE-bench Verified for coding, 99.6 on GSM8K for math, and 66.7 on GPQA for reasoning. The model supports English and Chinese, and it's tagged for agent capabilities. This is almost certainly a Mixture of Experts architecture. A dense 1T model would need terabytes of VRAM, making it impractical for open-source deployment. With MoE, only a fraction of parameters activate per token. That's why Gert Labs found MiMo cheaper to run than Moonshot AI's Kimi K2.6 while delivering slightly better tool use. **Mario Zechner** recently launched a **coding agent pi**. There's a wrinkle. The default system prompt declares MiMo a "Chinese AI model" that must comply with all applicable laws and regulations of the People's Republic of China. You can override the system prompt. But anyone evaluating this model should test it against their actual use case first, because those guardrails ship out of the box.
Xiaomi's 1T Open-Source Model Runs Cheaper Than Kimi K2.6
Xiaomi open-sourced MiMo-V2.5-Pro, a 1 trillion parameter language model with a 1M-token context window and MIT license. Benchmarks put it competitive with proprietary options: 66.7 on GPQA reasoning, 99.6 on GSM8K math, 78.9 on SWE-bench Verified coding. A Gert Labs tester found it comparable to Kimi K2.6 with slightly better tool use and lower inference cost. Supports English and Chinese, tagged for agent and long-context work.