GoKawiilResearcher builds deterministic RISC-V CPU inside Qwen 3.5 model weights
An independent researcher released 'Qwenputer,' a modified Qwen 3.5 4B model that embeds a functioning RISC-V CPU within its own weights, using the model's Gated DeltaNet attention layers as RAM. The CPU runs entirely inside the model's forward passes, executing arithmetic, C and MicroPython code without any external tool, plugin, or sandbox. The project, along with its code, is published on Hugging Face, ModelScope and GitHub, with two follow-up write-ups exploring related implications.
GoKawiil's interpretation of the reporting above, not reported fact.
The project suggests that linear-attention mechanisms like Gated DeltaNet can be repurposed as exact, addressable memory, blurring the line between a language model's statistical reasoning and literal computation. The researcher notes this could have double-edged implications: such embedded computation might let models hide intermediate reasoning in ways that evade chain-of-thought monitoring, but could also be used to build tamper-resistant internal monitoring tools. These remain exploratory, unverified directions rather than demonstrated production risks or safeguards.
- Qwenputer embeds a RISC-V CPU inside Qwen 3.5 4B's weights, using Gated DeltaNet attention states as RAM.
- It runs arithmetic and executes C/MicroPython code purely through the model's forward pass, without external tools.
- The researcher published follow-up analyses on potential risks (hidden reasoning evading monitoring) and benefits (tamper-proof in-model monitoring).
Source: divagr.com, 2026-10-29
Published there as: “Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.