Divyansh Agrawal has put a working computer inside a language model. In a blog post dated 22 October 2026, the developer describes Qwenputer: a RISC-V processor (an open chip design) written directly into the weights of Alibaba’s open Qwen 3.5 4B. When the model needs to add or multiply big numbers, it does not reach for a calculator or a Python sandbox. The sums run inside the model itself.
Language models are famously bad at exact arithmetic. They predict what a plausible answer looks like, and for long numbers plausible is often wrong. The usual fix is a tool: the model writes out the problem, an outside program solves it, and the answer comes back. Agrawal’s version removes the outside program. Every step of the calculation happens in the same pass of numbers through the network that produces the words.
The trick starts with memory. Qwen 3.5 builds most of its attention layers from a design called Gated DeltaNet, which keeps a running summary of the conversation in a grid of numbers. Agrawal found that if you pin two of the layer’s settings to 1 and address each row of the grid with a single marker, the grid behaves like ordinary computer memory. Write to row 12 and you get exactly what you wrote when you read row 12 back. That kind of exactness is normally the one thing these layers are not built for, and the grid never grows, no matter how long the chat gets.
With memory in place, no training was needed for the rest. A compiler converts a description of the processor into weights by hand, turning each small step, such as reading a register or jumping when a value is zero, into a piece of a network layer. The processor occupies 11 layers placed in front of the model’s original 32, and the original weights are untouched. As Qwen generates its answer, a few extra tokens per instruction act as a running log of what the processor is doing.
Agrawal reports that the chip passes the official RISC-V test suite and ran 18 million instructions on its own without an error. He also checked roughly 230,000 of its printed results against a reference machine, and every one agreed. These are his own checks, published alongside open code on GitHub and models on Hugging Face and ModelScope, so anyone can rerun them.
On arithmetic sets from DeepMind, Qwenputer scored 97.5 percent on large division, against 72.5 percent for the plain model and 95 percent for the plain model with a Python tool. On large multiplication it scored 97.5 percent, behind the Python tool’s 100 percent. On a long-integer test called NUPA, it went from 48.6 percent to 100 percent, level with the Python tool. So the built-in computer matches an outside calculator without needing one.
The costs are real. The checkpoint is 18.8 GB, roughly double the base model, though Agrawal says most of the extra is zero padding. It needs a small driver program to boot, so a plain generate call will not run it yet. And it is slow: each processor instruction costs a few generated tokens, and his batched setup reached roughly 3,200 processor tokens each second on an L4 graphics card. A large calculation takes seconds, not milliseconds.
Agrawal also built an experimental file system on the same memory. A model can save a file by name, and the state can be carried into a later conversation. In his test, Qwen got facts that had changed between chats right 90 percent of the time, against 30 percent when the old chats were simply pasted back in.
The larger point is control. A company that can write exact, fixed behavior into a model’s weights could add a checker the model cannot switch off, and Agrawal has written two follow-ups on safety monitoring along those lines. For teams paying to run a tool server beside every model, a calculator that lives in the weights shows where that bill could shrink once the speed catches up.
Reported by Divyansh Agrawal on his blog (divagr.com) on 22 October 2026.
