Alibaba Researchers Develop QwenPU Architecture for LLM-in-LLM Processing

Technical Development of QwenPU

A new technical demonstration, titled “QwenPU,” has been released by developer divagr18, showcasing a method for running a Large Language Model (LLM) within the weights of another LLM. According to documentation provided on GitHub and X (formerly Twitter), the project involves running Andrej Karpathy’s “stories260k” program on a Neural Processing Unit (NPU) that has been compiled directly into the weights of the Qwen 3.5-4B model.

The developer states that the “stories260k” program is a C program. To facilitate its execution, an NPU capable of performing matrix multiplications (matmuls), softmax, and attention operations with bit-exact precision was integrated into the Qwen 3.5-4B weights. Notably, the implementation reportedly requires no external tools, as the LLM weights reside within the model’s own memory.

Implementation Details and Performance

The project, which follows the developer’s previous work involving a RISC-V CPU integrated into Qwen, aims to achieve high execution speeds through specific optimizations. While the developer has indicated that a full technical write-up is currently in progress, the initial release includes the source code on GitHub and the corresponding weights on Hugging Face.

The developer noted that work is ongoing regarding size reductions and further optimizations. As of the current report, the claims regarding the efficiency and speed of this “recursive LLM” architecture remain unverified by independent third-party benchmarks.

Community and Technical Reception

The announcement has generated discussion within developer communities on Reddit, where users have characterized the development as a form of virtualization for local LLMs. While some users have questioned the practical utility of the project—asking whether it serves a specific use case or is primarily a proof-of-concept—others have noted the technical complexity of compiling hardware-level operations into model weights.

The project highlights an emerging area of research regarding the virtualization of computational processes within neural network architectures. However, without independent verification of the performance metrics or the stability of the bit-exact NPU, the practical application of QwenPU for general-purpose computing remains speculative.