An X post dated October 29, 2026, showed Doom E1M1 running on Qwenputer, an experimental version of Qwen 3.5 4B with a 32-bit RISC-V CPU compiled into its model weights. The demonstration puts a striking face on the project’s central idea: CPU instructions can execute during the model’s forward passes, using state inside the model as memory.
How Qwenputer runs a CPU inside the model
Qwenputer adds 11 compiler-generated CPU layers ahead of Qwen’s 32 layers. The core implements RV32IM, the Zba extension and a small unit for big-number arithmetic; the original Qwen weights remain unchanged. Divyansh Agrawal describes the design and its implementation in the Qwenputer project write-up.
The memory comes from Gated DeltaNet, an attention mechanism used in three out of every four Qwen 3.5 attention layers. Agrawal configures its state for addressable reads and writes: with the gate and write strength set to 1 and one-hot keys, individual state rows can serve as RAM.
CPU instructions travel as generated trace tokens and execute in the model’s forward passes. Qwen uses a tool-call-like format to communicate with the internal CPU, but that format does not hand the arithmetic to an external calculator. The current setup still needs the RvChat driver to load firmware and handle trace tokens; it does not run through a plain generate() call alone.
The key design choice is where the computation happens: inside the model’s forward passes, with RvChat still handling setup and trace tokens.
What the reported arithmetic scores compare
Agrawal reports accuracy scores for five named arithmetic tests, comparing Qwenputer with base Qwen 3.5 4B and Qwen 3.5 4B using a Python tool. The scores show Qwenputer ahead of the base model on all five listed tests. Against the Python-tool configuration, results vary by task: Qwenputer scores higher on division and the longer multiplication-and-division test, ties on addition or subtraction and NUPA, and scores lower on multiplication.
| Arithmetic test | Base Qwen 3.5 4B | Qwen 3.5 4B with Python tool | Qwenputer |
div_big | 72.5% | 95% | 97.5% |
mul_big | 55% | 100% | 97.5% |
mul_div_multiple_longer | 90% | 97.5% | 100% |
add_or_sub_big | 87.5% | 97.5% | 97.5% |
| NUPA | 48.6% | 100% | 100% |
These are Agrawal’s reported results for those specific test sets, not a general measure of model quality. He also reports that Qwenputer passed official riscv-tests, matched programs run in QEMU, and completed 18 million free-running instructions without an error. In a separate check, about 230,000 CPU output positions replayed against a reference machine reportedly had zero mismatches.
The setup trades speed for an unusual execution path
Agrawal says the 18.8 GB checkpoint runs on a single 24 GB GPU. Each CPU instruction costs several decoded tokens, so large calculations take seconds rather than milliseconds. For a batched benchmark executor on an L4, he reports throughput of about 3,200 CPU tokens per second. That is a token rate for that executor, not a general CPU instruction rate.
Qwenputer also has an experimental filesystem built on Gated DeltaNet state. It has 1,024 directory slots and can hold up to 512 files, or about 53,000 text tokens. Files carry into a later conversation when the state is saved and then reloaded.
