Current outcome
Qwenputer has implemented a working 32-bit RISC-V virtual CPU and exact read/write memory inside Qwen3.5-4B weights, can execute compiled C programs and pass RISC-V instruction tests, and scores 39/40 on large-number multiplication; however, the model grows to 18.8GB and is much slower than external tools.
Progress timeline
2 material updates- #01
Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5
A developer implemented a functioning, deterministic RISC-V CPU and filesystem inside Qwen3.5 by exploiting gated delta network layers as exact RAM-like storage.
Source evidence: csmantle
- #02
开发者把32位CPU编进千问权重：大模型内部直接运行计算机程序
The report adds implementation details and test results for Qwenputer: developer Divyansh Agrawal prepended 11 layers to Qwen3.5-4B's original 32 layers to encode a 32-bit RISC-V virtual CPU into the weights, using Gated DeltaNet for exact read/write memory; the CPU runs compiled C programs, executed about 18 million instructions without detected errors, and scored 39/40 on DeepMind's large-number multiplication test versus 22/40 for base Qwen. The approach is inefficient, with the model file growing to 18.8GB, about twice the original, due to many zero-valued parameters.
State after update: Qwenputer has implemented a working 32-bit RISC-V virtual CPU and exact read/write memory inside Qwen3.5-4B weights, can execute compiled C programs and pass RISC-V instruction tests, and scores 39/40 on large-number multiplication; however, the model grows to 18.8GB and is much slower than external tools.
Source evidence: theblockbeats