Computation embedded in LLM weights: the QwenPU case and what to audit in open models
The author describes a tiny LLM running on an NPU compiled into the weights of Qwen 3.5-4B, and a cited post reports a RISC-V CPU embedded in the same weights, using the Gated DeltaNet layers as memory. The case shows that weights can hold computation that is not evident.
By Wendelmaques Pereira · Updated today
Source: QwenPU: um LLM pequeno rodando dentro dos pesos do Qwen 3.5-4B (x.com). Text prepared with AI from this source.
What happened and what to do
The author's post describes QwenPU, a small LLM (stories260k) running on an NPU compiled inside the weights of Qwen 3.5-4B. A cited post describes Qwenputer, a RISC-V CPU embedded in the same weights, in which the Gated DeltaNet layers act as RAM. The material shows that computation can be compiled into a model's weights, which matters for anyone studying the internals of recent architectures.
For a company, the practical point is that an open-weight model may carry behavior that common tests do not reveal. An engineering response includes checking the provenance of weights before adoption, running behavior and stress test suites outside typical use, comparing model versions with hashes and change records, and monitoring production outputs with alerts for anomalous patterns. When the model runs on the company's own infrastructure, this monitoring can be built as a continuous collection and evaluation pipeline, without sending sensitive data to third parties.
How the consultancy can help
Diagnosis of your company's exposure to open-weight models: provenance, behavior testing and review of the architectures adopted. Then implementation of an evaluation and monitoring pipeline on your own infrastructure, with continuous operation.
Next step
Send a short description of your open-model use case and receive a scoped proposal from Wendelmaques.
Let's assess your project
I reply with scope and a proposal.