Qwenputer turns Gated DeltaNet state into exact RAM for a CPU (theainewsreport.com)

Qwenputer, a groundbreaking project by the developer behind divagr.com, has successfully integrated a 32-bit RISC-V CPU into the Qwen 3.5 4B AI model, utilizing the model's Gated DeltaNet layers as exact RAM. This innovative approach demonstrates how the architecture can perform arithmetic tasks with remarkable accuracy, completing 18 million instructions without error. By embedding additional layers that execute computations internally, Qwenputer eliminates the need for external tool calls, significantly streamlining the process of generating precise outputs within the model's framework. The significance of Qwenputer extends to the AI/ML community, particularly in the realms of mechanistic interpretability and the development of agents that operate without traditional computational runtimes. While it operates on a modest speed of approximately 3,200 trace tokens per second, resulting in several hundred instructions per second, its ability to compute exact answers without external support is a notable advancement. With potential applications across various domains, researchers and developers are encouraged to explore Qwenputer by utilizing available resources on GitHub and Hugging Face. This project highlights the evolving complexity and capabilities of AI models, paving the way for more integrated computational solutions.