Article · AI translation
IT Home reported on October 30 that developer Divyansh Agrawal released the experimental model Qwenputer on October 22. Built on Alibaba's Qwen3.5-4B, it implements a 32-bit RISC-V CPU and memory functions within the neural network, allowing the large language model to execute programs and perform exact calculations without calling external tools such as Python.
Citing a blog post, IT Home explained that traditional large AI models generate answers by predicting the next token. They are prone to errors with large-integer multiplication and division or long chains of calculations, and usually need to call an external calculator or code interpreter.
Qwenputer's approach is to “compile” the RISC-V RV32IM instruction set and related extensions into the model weights and use Qwen3.5's gated DeltaNet internal state as RAM, thereby building a CPU within the model that can perform arithmetic, conditional branching and program execution.
Test results show that, for large-integer multiplication, the original Qwen3.5-4B achieved 55% accuracy, while Qwenputer reached 97.5%; for large-integer division, accuracy rose from 72.5% to 97.5%; and on tests involving multiple multiplication and division operations, it reached 100%.
On the NUPA long-integer arithmetic benchmark, Qwenputer achieved 100%, compared with just 48.6% for the standard version. The developer said the CPU passed the official RISC-V tests and executed approximately 18 million consecutive instructions without an error.
The project also experimentally adds an “in-model file system” that can store up to 512 files, or about 53,000 tokens, in the gated DeltaNet state and load them in later conversations. In tests, answers using information saved from earlier conversations achieved 90% accuracy, compared with just 30% for a control method that fed the earlier conversation content directly into the model.
However, Qwenputer still comes with clear costs: the model has grown to about 18.8 GB, nearly twice the size of the original; each CPU instruction requires generating multiple tokens, so calculations that originally took milliseconds may take seconds; and running it requires a dedicated driver, meaning it cannot be loaded directly like an ordinary large language model.
The project is open source on Hugging Face and GitHub and can run on a single GPU with 24 GB of video memory.
References
Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5
Click here to visit divagr1925/Qwen3.5-4B-Qwenputer on Hugging Face
Click here to visit the GitHub project
Advertising disclosure: External links in this article, including but not limited to hyperlinks, QR codes and passcodes, are provided to convey more information and save time in selecting information. The results are for reference only. All IT Home articles containing external links include this disclosure.