Xiaomi MiMo-V2.5 ships 1M-context open weights

Xiaomi is no longer dabbling at the edge of the open-model race; it is shipping frontier-scale weights that look aimed at serious agent workloads, not hobbyist demos.

• –MiMo-V2.5-Pro is the heavier code and agent model, while MiMo-V2.5 is the more compelling generalist because native omni-modal support matters more for real workflows
• –MIT licensing lowers the friction for commercial use, fine-tuning, and self-hosted deployment, which is the real unlock for teams that want control
• –1M context is only useful if the serving stack is solid, so the practical story here is as much about infra compatibility as benchmark numbers
• –The open question is whether these MoE giants stay credible outside curated evals once developers push them through long tool chains and messy production prompts
• –If Xiaomi’s numbers hold up, this strengthens the open-weight frontier alongside DeepSeek and Qwen, especially for agentic and multimodal systems