Xiaomi MiMo-V2.5 Series Open-Sourced: 1T Parameters under MIT License, Token Efficiency Surpassing GPT-5.4 on ClawEval

According to monitoring by Dongcha Beating, the Xiaomi MiMo team has open-sourced the MiMo-V2.5 series of large models, which includes two models, both under the MIT license, supporting commercial deployment, continued training, and fine-tuning, with a context window of up to 1 million tokens. The MiMo-V2.5-Pro is a pure text MoE model (Mixture of Experts architecture) with a total of 1.02 trillion parameters and 42 billion active parameters; MiMo-V2.5 is a native multimodal model with a total of 310 billion parameters and 15 billion active parameters, supporting text, image, video, and audio understanding. MiMo-V2.5-Pro primarily targets complex agent and programming tasks. In the ClawEval evaluation, V2.5-Pro achieved a 64% Pass^3, reaching comparable levels while consuming only about 70,000 tokens per task trajectory, which is approximately 40% to 60% less than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4. The SWE-bench Verified score is 78.9. In a case showcased on the official blog, V2.5-Pro autonomously implemented a complete SysY to RISC-V compiler for a compiler principles project at Peking University, taking 4.3 hours and 672 tool calls, achieving a perfect score of 233/233 on a hidden test set. MiMo-V2.5 is designed for multimodal agent scenarios, equipped with a dedicated visual encoder (729 million parameters ViT) and an audio encoder (261 million parameters), scoring 62.3 on the Claw-Eval general subset. Both models utilize a mixed architecture of sliding window attention (SWA) and global attention (GA), along with a 3-layer multi-token prediction (MTP) module (predicting multiple tokens at once to accelerate inference). Weights have been released on Hugging Face. Along with the open-source release, the MiMo team has launched the ‘Orbit Trillion Token Creator Incentive Program’, offering a total of 100 trillion token quota free to global users within 30 days. Individual developers, teams, and enterprises can apply on the event page, with an evaluation period of about 3 working days. Upon approval, benefits will be credited in the form of Token Plan or grants, which can be directly used with programming tools like Claude Code and Cursor.