Running Qwen 3.8 Flash Next Locally on a 12GB Android Phone
A developer successfully ran the Qwen 3.8 Flash Next model, which has an 80GB footprint, locally on a mid-range 12GB Android phone. The model achieved a processing speed of 3.5 tokens per second thanks to custom optimizations and low quantization on its dense parts. This achievement demonstrates that massive large language models (LLMs) can be run on consumer-grade mobile hardware without relying on cloud servers. It opens up new possibilities for private, offline, and cost-effective edge computing applications on everyday smartphones. The setup runs on a $400–$500 mid-range phone by utilizing aggressive quantization on the dense parts of the model to fit the 80GB footprint into 12GB of RAM. The developer achieved a usable speed of 3.5 tokens per second, showing the viability of mobile-side LLM execution.
## BACKGROUND
Qwen is a family of large language models developed by Alibaba Cloud. Quantization is a model compression technique that reduces the precision of a model's weights, drastically lowering memory requirements and allowing large models to run on resource-constrained hardware like mobile devices.