Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5

Large language models can now execute code and store files internally, potentially reducing reliance on external tools and cloud infrastructure for certain tasks.

An experimental model named Qwenputer integrates a functioning RISC-V CPU directly into the weights of the Qwen 3.5 4B large language model. This CPU utilizes the model's own Gated DeltaNet attention layers as RAM, enabling it to perform arithmetic operations, and run C and MicroPython code without external tools. The RISC-V CPU, a 32-bit RV32IM core with extensions, occupies 11 layers of the model and has successfully executed 18 million instructions without errors. Qwenputer also includes an experimental in-model file storage system capable of holding up to 512 files. Performance benchmarks show Qwenputer matching or exceeding the accuracy of the base Qwen model when provided with external Python tools for arithmetic tasks, and demonstrating improved factual recall with its integrated filesystem compared to pasting previous chat logs.

This innovation demonstrates a novel approach to embedding computational capabilities directly within the architecture of large language models, moving beyond traditional tool-use paradigms. By repurposing the model's attention layers as RAM and integrating a RISC-V CPU, Qwenputer showcases a path towards more self-contained and efficient AI systems. The deterministic nature of the CPU and its ability to run code internally suggest significant implications for model reliability, security, and performance in applications requiring precise computation.

The development of Qwenputer signals a potential shift towards highly integrated AI agents capable of complex internal processing and persistent memory. Its success in matching external tool performance on arithmetic tasks and improving factual recall through its filesystem highlights the viability of such an approach. Future advancements may focus on optimizing the speed of these in-model computations and expanding their functionality, potentially blurring the lines between software execution and neural network processing.