early this morning, xiaomi technology officially announced that the mimo-v2.5 series models have been fully open-sourced, while simultaneously launching the “orbit 100 trillion token initiative.” the series entered public beta on april 23, comprising two variants—mimo-v2.5-pro and mimo-v2.5—both of which support a 1-million-token context window.
the pro version is designed for complex tasks, with deep integration for agent and coding applications, and currently ranks first globally among open-source models on the gdpval-aa and claweval leaderboards. the other variant is a native multimodal model that supports understanding of text, images, videos, and audio, boasting powerful agent capabilities. xiaomi emphasizes that the true value of these models lies not in leaderboard rankings, but in their ability to efficiently solve developers’ real-world problems. following extensive refinement during the public beta, the series now adheres to open standards: all model weights, including the base model, are fully released under the permissive mit license, allowing for free commercial use, retraining, and fine-tuning.
the concurrently launched mimo orbit program consists of two components: the “100 trillion token creator incentive program” for ai builders, and the “agent ecosystem co-creation program” for agent-framework teams. over a 30-day period (april 28 to may 28, beijing time), xiaomi will distribute a total of 100 trillion tokens to users worldwide, with the giveaway ending once all tokens have been distributed. users can apply by visiting 100t.xiaomimimo.com; approved applicants may receive up to the max-tier token plan, which includes 1.6 billion credits valued at rmb 659.
in addition, xiaomi is providing token-based incentives to agent-framework teams worldwide, enabling users to access and experience the mimo series models free of charge. currently, mimo-v2.5-pro has already achieved day-one compatibility with chip manufacturers such as alibaba’s t-head, amazon web services, amd, baidu’s kunlun chip, suiyuan technology, moxie, and tianshu zhixin, while also completing day-zero integration with mainstream inference frameworks sglang and vllm, ensuring that the open-source models are immediately available globally.
