Divyansh Agrawal describes Qwenputer, an experimental modification of Qwen3.5-4B that executes a RISC-V CPU inside the model's forward passes, in his October 22, 2026 write-up. The project puts calculation inside the modified model, but its RISC-V version still needs a driver to manage execution. The supplied results are the developer's tests; no independent measurements accompany this article.
A compiler supplies the computer
The repository describes 11 compiled transformer layers placed before Qwen's own layers. A compiler sets the computer's weights directly, while Qwen's original weights stay unchanged. This is a constructed machine inside a modified model. It does not demonstrate that the base model learned to execute RISC-V.
Memory comes from Gated DeltaNet, the recurrent state used by Qwen's architecture. The project configures it so a write replaces one addressed row and a read retrieves that row. The CPU stores its program and working data there. Generated trace tokens record the machine's execution, according to the repository's technical description.
Qwen requests a calculation through the tool-call syntax it was trained to use. The added layers execute the calculation and emit the response. The repository says the calculation happens internally: an external Python interpreter does not execute that request. The familiar request format and the location of execution are separate details.
Internal execution still needs a driver
The RISC-V version needs RvChat to load firmware into the DeltaNet state and keep CPU trace tokens out of Qwen's language layers. Agrawal's limitations section says it doesn't run with a plain generate() call yet. Developers evaluating this version should start with the repository's RISC-V quickstart and its driver requirement.
That distinction also matters when reading the repository: its separate, earlier CHIP-8 version has a plain generate() example. That example does not establish the same compatibility for the RISC-V version discussed here.
The repository's benchmark table reports greedy decoding and 40 problems per set. On synthetic integer arithmetic, Qwenputer scores 97.5%, versus 75.0% for base Qwen3.5-4B and 97.5% for base Qwen with a host Python tool. These are developer-reported arithmetic results. This small, targeted evaluation does not establish general reasoning gains or a speed advantage over Python.
There are costs to placing the CPU here. Agrawal's write-up lists an 18.8 GB checkpoint, mostly zero padding. He also reports that a big calculation takes seconds rather than milliseconds because each CPU instruction costs a few decoded tokens. Developers exploring compiled computation inside model weights have a concrete experiment to inspect. Those who need plain generate() compatibility for this RISC-V version should account for the driver limitation before integrating it. Independent replay of execution traces and measurements of complete calculation latency would help assess it beyond the published arithmetic scores.