A small AI model has a working computer chip inside it
A researcher built a real RISC-V processor out of a language model's own memory.
This shows how blurry the line between a model and a computer is getting, which matters for anyone trying to reason about what these systems can really do by themselves, with no outside tools. A researcher built a project called Qwenputer by taking a stock (unmodified) Qwen 3.5 model with 4 billion parameters (the model's internal settings) and adding one thing: a RISC-V chip, a common and simple processor design, built directly into the model's own weights. That chip runs entirely inside the model's own process of writing out an answer, and it can do arithmetic and run real C and MicroPython code with no outside tool, sandbox, or separate computer behind it at all.
The trick rests on a kind of layer called Gated DeltaNet (GDN), a lighter alternative to standard attention that three out of every four layers in Qwen 3.5 use to track context. Each GDN layer keeps a running state, like a notebook, that normally fades a little with every new word based on a 'gate' setting. The researcher found that by setting that gate to never forget, setting the write strength to fully overwrite instead of blend, and addressing memory with a method that points at exactly one slot, a GDN layer turns into exact, addressable RAM (working memory): write to one slot, read it back perfectly, and it costs far less than using the model's normal attention as memory. The underlying math was shown to work this way in an earlier paper; what is new is making it exact and addressable inside a real, working model, reliably enough to run a whole chip on top of it.
No training was involved in building the chip itself. A compiler takes a plain description of the chip, steps like 'read this slot', 'add these two numbers', or 'jump if zero', and turns each one directly into a slice of a transformer layer: attention heads that look values up, feed forward units that do the math, and GDN heads that hold the memory. The chip runs as the model generates text, producing a small burst of tokens (chunks of text, roughly three quarters of a word each) for every instruction it executes.
The chip itself is a 32 bit RISC-V core (the RV32IM instruction set, plus one extra math extension and a unit for big number arithmetic), and it takes up 11 of the model's 32 total layers. It cannot run Linux yet, but it covers the full instruction set and runs real C and MicroPython programs, just slowly. On a big number division test, Qwenputer scored 97.5 percent, beating the 95 percent a base Qwen model gets when handed a separate Python tool to do the math, and far ahead of the 72.5 percent base Qwen scores with no tool at all.