Qwenputer: claimed RISC-V CPU inside the Qwen3.5-4B weights

What happened and what to do

The post by @divagr1925 presents the Qwenputer. According to the author, a Qwen3.5-4B model has a RISC-V CPU encoded in its own weights. As described, the Gated DeltaNet layers act as RAM, allowing C and MicroPython code to run during forward passes. The author also shows the use of Gated DeltaNet state as general-purpose memory. The author's benchmark claims have not been verified.

For a company, the practical point is not to adopt the technique but to treat the claim as a hypothesis to test. A reproducible evaluation pipeline could run the cited checkpoint on private infrastructure, record inputs, outputs, execution time and resource use, and compare the results against baselines, with versioning of weights and prompts. In parallel, provenance monitoring is worthwhile: verify model file hashes, origin and licenses before any production use. That way the decision rests on measurements, not isolated demonstrations.

How the consultancy can help

Model evaluation diagnosis: we define hypotheses, criteria and a reproducible test set to verify performance and behavior claims, plus an analysis of weight provenance. We then implement the evaluation harness on private infrastructure, with metric logging and reports.

Next step

If your team evaluates open models before adopting them, send a short description of the case and your decision criteria. We will reply with a scoped diagnosis and implementation proposal.