Qwen puter embeds a 32-bit RISC-V virtual CPU into Qwen the weights and passes approximately 18000000 instruction tests

AT A GLANCE

Qwen puter embeds 32-bit RISC-V virtual CPU into Qwen the model weights, enabling programs to run and calculations to be performed within the model.

Article

Developers have 32-bit CPU compiled into Qwen's weights: large language models directly run computer programs internally

Beating AI News reports, independent developer Divyansh Agrawal built Qwen3.5-4B based on Alibaba's Qwen puter, embedding a 32-bit RISC-V virtual CPU directly into the model weights, allowing large language models to perform precise calculations while generating responses without calling external computing tools.

The author did not retrain Qwen; instead, 32 additional layers were added before the original 11-layer network, using a compiler to write the operating rules of CPU directly into the model parameters. The added network also uses a Gated DeltaNet structure to implement precisely readable and writable memory for storing programs and data.

This virtual CPU can already run compiled C programs and passed RISC-V instruction tests, with no errors found in tests involving approximately 18000000 instructions in total.

In DeepMind's 40 large-number multiplication problems, the original Qwen answered 22 correctly, while Qwen puter answered 39; after the original Qwen was connected to external Python, it answered all of them correctly.CPU itself can execute instructions accurately, but the large language model may still make errors when invoking it or outputting results.

This technical approach has precedents. This year,March Percepta used a neural network to implement a virtual computer capable of running C programs. However, it was a purpose-built model.Qwen puter instead integrates similar computing capabilities into an already trained Qwen so that, while generating responses, the language model can also use its internal CPU to perform calculations.

However, this solution is currently very inefficient. A large number of zero-valued parameters expand the model file to 18.8GB—approximately twice the size of the original—and the computation speed is also far lower than that of external tools.

Original link https://m.theblockbeats.info/flash/374349

Key points

Independent developer Divyansh Agrawal built Qwen3.5-4B based on Alibaba’s Qwen puter。
Qwen puter embeds a 32-bit RISC-V virtual CPU machine into the model weights.
The authors added 32 layers before the original network and added 11 layers, writing the CPU computational rules as model parameters.
The added network uses Gated DeltaNet its architecture to implement readable and writable memory for storing programs and data.
The virtual CPU machine can now run compiled C programs and pass RISC-V instruction tests.
In tests involving approximately 18000000 instructions executed cumulatively, no errors were found.
In DeepMind's 40 large-number multiplication problems, the original Qwen answered 22 correctly, while Qwen puter answered 39 correctly.
Original version Qwen After connecting external Python, 40 questions were answered correctly.
The model file reaches 18. 8GB, approximately twice the size of the original version, with a calculation speed far inferior to that of external tools.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Typically, large models need to call external calculators or Python for precise calculations. This project integrates a virtual CPU computer and memory into the model itself, enabling it to run compiled C programs; however, when the model calls CPU or outputs results, errors may still occur.

Why it matters to readers

It demonstrates that a trained language model can acquire internal program execution capabilities by adding network layers and parameters, while also revealing deployment limitations in terms of model size and speed.

It can be understood as equipping a large model with a small virtual computer, enabling it to execute programs internally rather than calling external tools each time.

Risks and unknowns

The core performance data reported mainly comes from the project developers; the evidence does not indicate that it has undergone independent third-party verification.
The evidence does not indicate whether the project code, model weights, and test scripts have been made publicly available and are reproducible.
The evidence does not provide the specific error rate when the model calls CPU and outputs calculation results.
The evidence did not specify whether subsequent measures would reduce 18. 8GB model size or increase computational speed.

Qwen puter embeds a 32-bit RISC-V virtual CPU into Qwen the weights and passes approximately 18000000 instruction tests

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center