X discussion: an indie developer compiles an NPU into Qwen3.5-4B weights and runs Karpathy's stories260K, an LLM inside an LLM

Background and the problem

Over the past two weeks, indie developer div (@divagr1925 on X) has posted two experiments that challenge the assumption that tools such as a Python interpreter must live outside the model. Qwenputer, on Oct 28, compiled a full 32-bit RISC-V CPU into the weights of Qwen3.5-4B. QwenPU, on Oct 31, adds a vector unit so that another language model runs inside the model. The post drew about 1,540 likes and 85K views and was collected into an X trending story. To be clear, this is a personal experiment, not an official release from Alibaba's Qwen team.

Architecture and how it works

According to the project README, QwenPU places 13 compiled computer layers, with registers and memory held in Gated DeltaNet state, in front of Qwen's 32 layers. They implement RV32IM, the Zba extension, a bignum unit and the new vector unit, widening the hidden size from 2560 to 5120 while leaving Qwen's own weights unchanged. No new weight is trained; a compiler computes every one. The int4 weights of stories260K are loaded as data into the unit's weight memory rather than baked into the checkpoint.

Benchmarks and measured numbers

All numbers below are author-reported, and no independent reproduction has been published. Trace tokens needed per story token are 2.43M for plain C, 225K with Qwenputer's NPU instructions, and 53K with QwenPU's vector unit. The GPU executor runs about 3,300 trace tokens per second on an L4, so one story token takes about 16 seconds. For the earlier Qwenputer on DeepMind math problems, the author reports 55-90% for the base model, 95-100% for the base model with a Python tool, and 97.5-100% for Qwenputer.

Getting started

Code is on GitHub at divagr18/QwenPU (MIT) and weights are on Hugging Face at divagr1925/QwenPU (Apache-2.0). Following the README setup, the story demo and checks that run only the computer layers fit a 24 GB GPU, while loading the full 26.7 GB checkpoint needs 40 GB or more. Discussion on X splits two ways: some say it shows model weights can do exact computation and that whether tools must be external is worth rethinking; others point out that it is extremely slow and that the checkpoint is mostly zero padding, so its practical value is limited. The author says a full write-up is still in progress, along with size reductions.