IT Home On October 30, developer Divyansh Agrawal released an experimental model Qwenputer on October 22. Built based on Alibaba’s Qwen3.5-4B, it implements 32-bit RISC-V CPU and memory functions within the neural network, allowing large language models to execute programs and perform precise calculations without calling external tools such as Python.
According to a blog post cited by IT Home, traditional large AI models generate responses by predicting the next token. They are prone to errors when dealing with large integer multiplications and divisions, or long chain operations, and usually require the use of external calculators or code interpreters.
And Qwenputer’s approach is to “compile” the RISC-V RV32IM instruction set and related extensions into the model weights, and use the Qwen3.5 Gate Control DeltaNet internal state as RAM, thereby building a CPU in the model that can perform executable arithmetic, conditional branches, and programs.
The test results show that in large integer multiplication, the original Qwen3.5-4B has a accuracy of 55%, while Qwenputer increases to 97.5%; large integer division improves from 72.5% to 97.5%; tests involving multiple multiplications and divisions reach 100%.
On the long integer operation benchmark NUPA, Qwenputer reached 100%, while the standard version was only 48.6%. Developers said that the CPU has passed the official RISC-V tests, and no errors occurred when executing approximately 18 million instructions in a row.
The project also experimentally added an “internal model file system” that can store up to 512 files and approximately 53,000 tokens in the Gate Control DeltaNet state, and load them during subsequent conversations. During testing, the accuracy of responses using information saved from previous conversations reached 90%, while the comparison method of directly inputting previous conversation content into the model was only 30%.
However, Qwenputer still has significant costs: the model size increases to approximately 18.8GB, nearly twice that of the original version; multiple tokens are generated for each CPU instruction, and calculations that took just a few milliseconds may now take several seconds; running it also requires specialized drivers, and it cannot be loaded directly like ordinary large models.
The project has been opened sourced on Hugging Face and GitHub, and can be run on a single GPU with 24GB of memory.
Reference