AI Engineer Compiles Working RISC-V CPU Directly Into Qwen Model Weights

A software developer has achieved a remarkable feat in artificial intelligence and computer architecture. Divyansh Agrawal successfully compiled a fully functional 32-bit RISC-V central processing unit directly into the weights of a large language model.

his experimental project, dubbed Qwenputer, proves that an AI model can operate as a standalone computer without relying on external plugins or traditional system memory.

The project utilizes the stock Qwen 3.5 4B model as its foundation. This specific model is part of a broader series of open-weights language models developed by Alibaba Cloud. Instead of training new weights through conventional machine learning methods, the developer inserted 11 compiled transformer layers directly in front of the model's existing 32 layers.

These specific layers act as the processor logic. The weights within these layers were not learned during training but were mathematically compiled to execute processor arithmetic operations.

To handle memory operations, the system repurposes the model's built-in Gated DeltaNet layers. This linear-attention mechanism, originally designed for long-context processing, effectively doubles as random access memory for the embedded CPU. The original 32 layers of the base model remain completely untouched. This allows the system to maintain its foundational language capabilities while simultaneously executing low-level machine code.

This architectural innovation enables the model to process actual software workloads. The compiled CPU successfully runs C code and MicroPython scripts by treating its own internal state as executable memory. In a classic demonstration of computing prowess, the system even rendered a frame of the legendary game Doom.

This experiment highlights the dual-use potential of modern neural networks. While not intended for commercial hardware replacement, Qwenputer demonstrates that the boundary between software weights and hardware logic is becoming increasingly blurred. Researchers can now explore how large language models might internally simulate complex computational environments.

The success of this compilation opens new avenues for AI safety research and model interpretability. Understanding how a neural network allocates its internal state to mimic a traditional CPU provides valuable insights into the hidden mechanics of transformer architectures. As open-source communities continue to push these boundaries, we will likely see more innovative intersections between artificial intelligence and classical computer engineering.