Xiaomi Releases Open-Weights MiMo-V2.6 Distillation Model Based on Qwen 9B
Xiaomi has released MiMo-V2.6-Distill-Qwen-9B on Hugging Face, an open-weights language model created by distilling knowledge from their MiMo-V2.6 architecture into a 9-billion parameter Qwen base. Distilling large foundation models into compact 9B architectures makes advanced capabilities accessible to local LLM users and resource-constrained hardware. It underscores a growing industry trend where leading tech companies share distilled open weights to foster local AI adoption. The release uses Alibaba's popular Qwen 9B model structure as the student network, trained on outputs from Xiaomi's flagship MiMo-V2.6 teacher model. The open-weights model is publicly available for download and local inference via the XiaomiMiMo organization repository on Hugging Face.
## BACKGROUND
Knowledge distillation in LLMs is a compression technique where a smaller 'student' model is trained to replicate the performance and reasoning of a much larger 'teacher' model. Xiaomi's MiMo-V2.6 family includes large models such as the 1T+ parameter MiMo-V2.6-Pro and the 309B MiMo-V2.6-Flash. Qwen is a family of widely used open-weights base models developed by Alibaba, known for strong reasoning and coding benchmark scores.