Xiaomi MiMo-V2.5 Series Open Source: 1T Parameters MIT License, ClawEval surpasses GPT-5.4 in token efficiency

According to Beating Monitoring, Xiaomi’s MiMo team has open-sourced the MiMo-V2.5 series large models, including two models, both licensed under MIT, supporting commercial deployment, continued training, and fine-tuning, with a context window of up to 1 million tokens. Among them, 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 310 billion total parameters and 15 billion active parameters, supporting text, image, video, and audio understanding.

MiMo-V2.5-Pro mainly targets complex agent and programming tasks. In the ClawEval evaluation, V2.5-Pro achieved a comparable level with a 64% Pass^3, but each task trajectory only consumed about 70k tokens, approximately 40% to 60% less than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4. It scored 78.9 on SWE-bench Verified. In the case studies showcased on the official blog, V2.5-Pro independently developed a complete SysY to RISC-V compiler for Peking University’s compiler principles course project, taking 4.3 hours and 672 tool calls, with a perfect score of 233/233 on the hidden test set.

MiMo-V2.5 is designed for multimodal agent scenarios. The model is equipped with a dedicated visual encoder (1M parameters ViT) and an audio encoder (1.02T parameters), scoring 62.3 on the Claw-Eval general subset. Both models adopt a hybrid architecture of sliding window attention (SWA) and global attention (GA), combined with a 3-layer multi-token prediction (MTP) module (predicting multiple tokens at once to accelerate inference). The weights have been released on Hugging Face.

Along with the open-source release, the MiMo team has simultaneously launched the “Orbit Trillion Token Creator Incentive Program,” offering a total of 1 trillion tokens free to global users within 30 days. Individual developers, teams, and enterprises can submit applications on the event page; the evaluation cycle is about 3 working days. After approval, benefits are credited in the form of Token Plans or grants, which can be directly used with programming tools like Claude Code, Cursor, and others.