Qwenputer is an experimental Qwen3.5-4B derivative by builder divagr (identified in secondary results as Divyansh Agrawal). The builder presents it as embedding a 32-bit RISC-V CPU in model weights, using Gated DeltaNet state as RAM; linked repositories, model weights and a demo are cited, and a separate Reddit post reports a slow Doom demonstration. On Nov. 1, a newer post quoting the builder’s X thread announced QwenPU: a neural processing unit compiled into Qwen’s weights to run the C program stories260k, with the builder claiming bit-exact matrix multiplication, softmax and attention and no external tools. The full write-up is still pending. The supplied evidence does not independently establish correctness, reliability or performance; much of the technical detail is in attributed claims, summaries and discussion.
The striking idea is to make a model’s recurrent state perform exact, structured computation rather than merely have the model describe or call an external computer. If dependable, this could enable unusual forms of self-contained computation and persistent state inside inference. For now, the demonstrations are experimental, and no evidence here shows an advantage over ordinary code execution or tool use. The newer QwenPU claim extends the idea from CPU emulation toward running a small neural network internally, but its account and optimizations remain forthcoming.
Qwenputer is a Qwen3.5-4B derivative presented by its builder as implementing a 32-bit RISC-V CPU and filesystem using model computation and state; a separate project reportedly ran Doom on its virtual machine, slowly. The builder later announced QwenPU, linking code and weights and claiming an NPU compiled into Qwen’s weights can run Karpathy’s stories260k program with bit-exact matrix multiplication, softmax and attention, without external tools. A separate post describes Qwen interpreting model weights, but it does not establish that Qwenputer or QwenPU works reliably. These remain experimental project claims; correctness, supported workloads, speed and practical advantage over ordinary execution remain open. The latest attention increase is on a distinct, adjacent post rather than new validation of this case.
now 0 pts/hpeak 308 pts/hcomments 0/hpeers p25momentum: steady3 platformsage 312h
points/hour across evidence · reading as of 2026-11-10 23:26:04.141941+11:00 · deterministic, not a model opinion
2026-11-07T19:28:21Z
A separate Qwen-weight interpretation post is drawing renewed attention, but it does not independently validate Qwenputer’s CPU or QwenPU’s NPU execution claims. The case remains a consequential technical demonstration with unresolved questions about correctness, reliability and practical utility; the attention spike is not evidence of progress on those questions.
2026-11-04T00:42:35Z
The QwenPU and QwenBend reports extend the experiment from CPU emulation to claims of running neural-network and typed functional workloads within model inference, but the latest related discussion adds no independent validation of those claims. Attention has cooled after the initial surge; the work remains consequential as an unusual systems demonstration, not evidence of a practical alternative to conventional execution.
2026-11-01T21:26:48Z
The newer QwenPU announcement extends the experiment from an embedded CPU to a claimed NPU running a small language-model program, and a separate QwenBend post describes another approach to computation within the model’s forward pass. These are meaningful architectural demonstrations to track, but the available accounts do not independently verify correctness or show practical performance; attention is now cooling from its earlier peak.
2026-10-31T19:26:08Z
audience grounding: medium — The striking idea is to make a model’s recurrent state perform exact, structured computation rather than merely have the model describe or call an external computer. If dependable, this could enable unusual forms of self-contained computation and persistent state inside inference. For now, the demonstrations are experimental, and no evidence here shows an advantage over ordinary code execution or tool use. The newer QwenPU claim extends the idea from CPU emulation toward running a small neural network internally, but its account and optimizations remain forthcoming.
2026-10-31T19:25:49Z
A new, thinly sourced report attributes a further step to the builder: running Karpathy’s stories260k inside an NPU compiled into Qwen’s weights. If confirmed, this extends the project from a CPU-emulation demonstration toward nested model execution, but the post points to the builder’s code, weights and an original social post rather than providing independent validation. Attention has otherwise cooled sharply from its earlier peak.
2026-10-30T05:29:38Z
The model-and-VM experiment remains a consequential demonstration, and the Doom follow-up shows a concrete use, but the newest HN item adds no distinct evidence. Attention has cooled from its peak; the main unresolved issue remains how much reliable execution is attributable to the model versus the surrounding executor.
2026-10-29T01:27:58Z
A separate builder reports running a basic, slow version of Doom on the Qwenputer virtual machine, adding a concrete demonstration that the system can support a nontrivial workload. This strengthens the implementation evidence, but does not establish independent reliability or practical advantage; attention is cooling from its peak despite a high relative engagement rate.
2026-10-28T23:29:09Z
A follow-on builder reports running a basic, slow version of Doom on the Qwenputer virtual machine, adding a concrete demonstration beyond the original CPU claim. This strengthens the case that the setup can execute at least some real workloads, but does not establish reliability, speed, or an advantage over ordinary external execution; attention is cooling from its peak.
2026-10-28T18:34:41Z
Qwenputer has moved from an unusual builder claim to a consequential, publicly available experiment: the project links code, weights and a live demo, while a second platform is now pointing to it. The technical distinction remains important—describing a CPU in model weights does not establish that execution is reliable or occurs without an executor—so its broader implications are still unsettled.
2026-10-28T13:32:46Z
audience grounding: medium — Qwenputer is an unusual experiment in embedding a defined computer-like execution mechanism inside a language model, rather than relying only on the model to imitate calculation or call an external tool. If the builder’s description holds, it could offer a way to combine language-model behavior with more predictable, bounded computation. For now, its significance is primarily technical and exploratory: the supplied sources do not establish broad usefulness, reliability, or an advantage over ordinary tool use.