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

According to DynaInsight Beating monitoring, the Xiaomi MiMo team has open-sourced the MiMo-V2.5 series large models, which include 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 1.02T parameters and 420 billion activation parameters; the MiMo-V2.5 is a native multimodal model with a total of 310 billion parameters and 15 billion activation parameters, supporting text, image, video, and audio understanding.

The MiMo-V2.5-Pro focuses on complex agent and programming tasks. In the ClawEval evaluation, V2.5-Pro achieved a 64% Pass^3 at the same level, but each task trajectory consumed only about 70,000 tokens, roughly 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 study showcased on the official blog, V2.5-Pro autonomously implemented a complete SysY to RISC-V compiler for a Peking University compiler principles project, taking 4.3 hours, 672 rounds of tool invocations, achieving a perfect score of 233/233 on the hidden test set.

The MiMo-V2.5 is designed for multimodal agent scenarios. The model comes with a dedicated visual encoder (7.29 billion parameters ViT) and audio encoder (2.61 billion parameters), scoring 62.3 on the Claw-Eval general subset. Both models feature a Sliding Window Attention (SWA) and Global Attention (GA) hybrid architecture, paired 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.

In conjunction with the open-source release, the MiMo team has simultaneously launched the "Orbit Trillion Token Creator Incentive Plan," distributing a total of 100 trillion tokens globally for free to users within 30 days. Individual developers, teams, and enterprises can submit applications on the event page, with an evaluation period of about 3 working days. Upon approval, benefits will be credited in the form of a Token Plan or a grant, which can be directly used with programming tools such as Claude Code and Cursor.