Developer Divyansh Agrawal has compiled a vector unit into the weights of Alibaba’s open Qwen3.5-4B model and used it to run Andrej Karpathy’s stories260K, a 260K-parameter Llama, inside the model’s own forward passes. Agrawal’s QwenPU repository, created on 31 October, says the build needs 53,000 trace tokens per story token.
A plain C build of the same story program needs 2,430,000 trace tokens per story token on the same in-weights CPU, so QwenPU takes 46 times fewer steps, by our arithmetic. It builds on Qwenputer, which the README describes as Qwen3.5-4B with “a 32-bit RISC-V CPU compiled and inserted directly into its weights”.
Short explainer · CPU Inside Qwen’s Weights: Karpathy’s LLM in 46x Fewer Steps
A vertical news Short summarising this story. Figures come from the primary sources cited below; our arithmetic is labelled.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how an AI model and the harness around it split the work. It teaches the general idea only and does not discuss Qwenputer or QwenPU.
The key insight: QwenPU runs Karpathy’s stories260K inside Qwen3.5-4B’s weights at 53,000 trace tokens per story token, 46 times fewer than a plain C build, by our arithmetic. On arithmetic, the README’s in-weights CPU averaged 98.0% against 97.5% for a Python tool, by our arithmetic.
The README says the CPU is “11 transformer layers in front of Qwen’s own”, and “none of their weights are learned: a compiler computes every one”. Qwen’s own weights are unchanged, it says.
The CPU uses the model’s Gated DeltaNet state as its RAM, keeping “its code, data, registers and program counter there”. It runs C and MicroPython code inside the forward passes and “passes the official riscv-tests suites”, the README says.
Qwen writes a tool call for a calculator in the format it was trained on, but “nothing outside the model ever runs it”. The README says “No runtime, sandbox or tool executor is involved”; a driver loads the CPU’s firmware into the DeltaNet state after the prompt.
Left: arithmetic accuracy of base Qwen3.5-4B, Qwenputer and base Qwen with a Python tool on five sets of 40 problems, from the Qwenputer README. Right: trace tokens per story token for three builds of Karpathy’s stories260K on the in-weights CPU. Means and the 46x ratio are FourWeekMBA arithmetic.
What Was Posted on X
1/3
A CPU inside an LLM is one thing, but running another LLM inside an LLM would be taking things a bit too far, wouldn’t it
Anyways, here’s an LLM running inside an LLM : QwenPU, an LLM (@karpathy’s stories260k) running on an NPU compiled inside Qwen 3.5-4B’s weights.
Developer Divyansh Agrawal (@divagr1925) introduced QwenPU on 31 October: Karpathy’s stories260K running on an NPU compiled inside Qwen3.5-4B’s weights. The piece takes its figures from the project’s GitHub READMEs.
1/4
Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model’s own Gated DeltaNet layers as RAM.
It runs C and MicroPython in its forward passes, no external runtime.
On DeepMind math: base 55โ90%, base + Python tool 95โ100%, Qwenputer 97.5โ100%.
Agrawal’s 28 October post introducing Qwenputer, Qwen3.5-4B with a RISC-V CPU inside its weights, with the DeepMind arithmetic ranges the piece checks against the Qwenputer README.
What the Arithmetic Tests Show
The README reports greedy decoding on 40 problems per set, comparing base Qwen3.5-4B, Qwenputer and base Qwen calling a real Python interpreter on the host. On DeepMind’s mul_big set, base scored 55.0%, Qwenputer 97.5% and the Python tool 100%.
Across the five sets, the averages were 76.0% for base, 98.0% for Qwenputer and 97.5% for the Python tool, by our arithmetic. The author says the project replayed about 230,000 of the computer’s output positions through a reference model and “found zero mismatches”.
The README for an earlier CHIP-8 version says word-problem benchmarks GSM8K and GSM-Hard were ties, “because word problems are limited by reasoning rather than arithmetic”.
The Cost: Seconds per Story Token
The README says the CPU runs at about 3,200 trace tokens a second on an L4 GPU for Qwenputer. On the earlier NPU build of stories260K, a fused executor at about 5,500 trace tokens a second means “a story token takes about 41 seconds”.
QwenPU cuts that to “about 16 seconds each on an L4”, the Qwenputer README says. The QwenPU README calls the project a “Work in progress”: its checkpoint is 26.7 GB, so loading it whole “needs a 40 GB or larger GPU”.
The Business Lever: Tools Moving Into Weights
The README contrasts Qwenputer with an external tool call: the same tool-call format is answered from inside the weights, and the model itself writes the response “digit by digit”. It says “It isn’t a plugin, a tool or a sandbox”.
The author’s 28 October post introducing Qwenputer had about 685,000 views on X when we read it. The Qwenputer repository was created on 19 October and had 362 stars on GitHub; its code is MIT-licensed and the weights Apache-2.0, as a modified Qwen3.5-4B.
The Structural Read
As we read the README, the project answers a tool call from inside the model: Qwen writes the call, and compiled layers run it and write the result.
As we read the speed figures, the cost moves from a sandbox to the model’s own tokens: the README gives about 16 seconds per story token on an L4 for QwenPU.
As we read the benchmarks, the gains are in exact arithmetic; the README says word problems were ties for the earlier CHIP-8 version.
Qwenputer README
“No runtime, sandbox or tool executor is involved”
Three Implications
FOR AGENT BUILDERS The README describes exact arithmetic answered inside the model’s weights, in Qwen’s existing tool-call format.
FOR MODEL HOSTS The QwenPU README says its checkpoint is 26.7 GB and needs a 40 GB or larger GPU to load whole.
FOR BUYERS All results are the author’s own, on 40 problems per set, with no independent replication in the pages read.
The framework says “Orchestration, memory, tool use, and workflow execution have become the new sources of competitive advantage.”
As we read it, Qwenputer moves one of those layers, tool use, from the harness into the model’s weights. The framework is the owner’s argument and does not discuss Qwenputer.
All results are the author’s own measurements, on 40 problems per set. The pages read report no independent replication or peer review. The pages read describe an individual developer’s project and do not name Alibaba as a participant. We did not run the models.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
A developer has compiled a RISC-V CPU and a vector unit into Qwen3.5-4B’s weights. The author’s tests show in-weights arithmetic matching a Python tool, and Karpathy’s stories260K running inside the model at about 16 seconds per story token.
A note on sourcing. All facts and figures come from the Qwenputer and QwenPU GitHub READMEs and repository data and the author’s X posts, read on 31 October 2026. Results are the author’s own measurements; means and ratios are our arithmetic. We did not run the models or contact the author.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
