A developer has published Qwenputer, a project that implements a working RISC-V CPU inside Qwen3.5. The design is described as deterministic, meaning identical inputs produce identical outputs rather than the variability usually associated with language models.
The project drew hundreds of points and more than a hundred comments on Hacker News, where readers debated how much of the behavior reflects genuine computation versus pattern matching learned during training. Supporters argued the deterministic output is the key detail, since a processor that produces unpredictable results would be useless.
Qwenputer joins a small but growing set of experiments that treat large models as substrates for structured systems rather than as conversational assistants. The appeal lies in testing the boundaries of what can be encoded and executed reliably inside a neural network.
Skeptics remain. Some commenters questioned whether the implementation generalizes beyond the specific programs tested, and whether the approach offers any practical advantage over a conventional emulator. For now the project stands as a demonstration, but one that has sharpened a broader argument about the nature of computation inside modern models.
The debate on Hacker News reflected a wider uncertainty in the field about what it means for a model to run a processor at all. A conventional emulator executes instructions through explicitly coded logic, while Qwenputer relies on whatever internal structure Qwen3.5 developed during training. That distinction sits at the center of the disagreement: if the CPU behavior emerges from learned patterns rather than designed circuitry, the question becomes whether the result counts as computation in any meaningful sense, or merely as a very convincing reproduction of one.
For the developer behind the project, the deterministic property appears to be the central claim. A language model that returns different answers to the same prompt would be unusable as a processor, since programs depend on repeatable results. Qwenputer's supporters on Hacker News seized on this point, arguing that consistency under repeated input separates the project from ordinary chatbot demonstrations.
Others were less convinced. Commenters asked whether the implementation holds up beyond the specific programs that were tested, and whether running a CPU inside a model offers any advantage over a conventional emulator, which is faster, cheaper and far better understood. Those questions remain unanswered for now, and the project stands as a demonstration rather than a practical tool. Even so, it has added a concrete example to a broader argument about the nature of computation inside modern models, and about how far structured systems can be pushed into neural networks before the analogy breaks down.