Qwenputer: A Functioning, Deterministic RISC-V CPU Inside Qwen3.5
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Developer Divyansh Agrawal reports building Qwenputer, an experimental version of Qwen3.5 4B with a compiled 32-bit RISC-V CPU that runs through the model’s forward passes. The system uses Gated DeltaNet state as addressable memory; Agrawal says it passes official RISC-V tests and completed an 18-million-instruction free-running test without error. It is a research demonstration, not a general-purpose computer or a feature of stock Qwen3.5.

Developer Divyansh Agrawal has built Qwenputer, an experimental version of Qwen3.5 4B that runs a 32-bit RISC-V CPU inside the model’s forward passes, without calling an external execution tool. The project uses the model’s Gated DeltaNet state as addressable memory and, Agrawal says, has passed official RISC-V tests and run 18 million instructions without an error.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

Qwenputer starts with stock Qwen3.5 4B weights and adds computer-specific weights compiled directly into 11 layers placed before the model’s original 32 layers. Agrawal says the CPU is an RV32IM core, with the Zba extension and a small bignum unit. The model performs CPU operations during decoding: generated trace tokens identify memory addresses, values and instruction steps. Those tokens represent execution inside the model rather than a request to an outside program.

The memory mechanism relies on Gated DeltaNet (GDN), a recurrent component used by most of the attention layers in the underlying model. Agrawal describes configuring the state update so a one-hot key selects a particular row, a write replaces the stored value, and a matching query retrieves it. That creates a form of exact, addressable storage for the CPU’s registers, memory and program counter. The report says a small driver, RvChat, loads firmware into the GDN state and keeps execution-trace tokens out of Qwen’s layers.

According to Agrawal, the core passes the official RISC-V test suite, its program results match QEMU, and a free-running test completed 18 million instructions without a single error. Free-running here means the CPU continued from its own outputs rather than being corrected by a reference machine at each step. The report also says it can execute C and MicroPython, though slowly. Agrawal characterizes it as an experiment, not a Linux-class processor.

At a glance
reportWhen: Reported in Agrawal’s blog; the source…
The developmentDivyansh Agrawal published an experimental Qwen3.5 4B model with a deterministic RISC-V CPU compiled into additional model layers.

A CPU Runs Through Model State

The experiment offers a concrete demonstration of a model’s internal recurrent state being used for more than ordinary conversational context. Here, GDN state acts as working memory, while compiled layers perform operations that produce repeatable CPU behavior. That is distinct from asking a language model to imitate calculations in text or handing code to a conventional software tool.

For researchers, the result is relevant to work on recurrent neural networks, memory and computation: the report describes turning a memory-like mechanism into storage with explicit addresses, then using it as part of a CPU. It also illustrates a possible way to combine learned model behavior with fixed, compiled components. The performance and reliability claims are Agrawal’s own project results, however; the supplied report does not describe an independent evaluation.

The distinction matters for practical use. Qwenputer is not evidence that an ordinary Qwen3.5 model automatically contains a working CPU, nor that the model can replace a standalone processor. It is a modified experimental build, and its reported ability to run code comes with the stated limitation that execution is slow.

Products Worth Considering

Sony WH-1000XM6 Wireless Noise-Cancelling Headphones – Midnight Blue

Sony WH-1000XM6 Wireless Noise-Cancelling Headphones – Midnight Blue

Samsung Odyssey OLED G8 34-Inch Curved Ultra WQHD Gaming Monitor

Samsung Odyssey OLED G8 34-Inch Curved Ultra WQHD Gaming Monitor

As an affiliate, we earn on qualifying purchases.

From Delta Memory to CPU

Gated DeltaNet is a recurrent attention mechanism in which each head maintains a state matrix and updates it as tokens are processed. Agrawal’s account says the model can use that state as memory by setting the gate to preserve stored values, using full-strength writes, and selecting rows with one-hot keys. In the experiment, the state matrix serves as RAM rather than simply carrying information forward through a conversation.

The underlying memory idea has earlier research roots. Agrawal points to a 2021 paper by Imanol Schlag, Kazuki Irie and Jürgen Schmidhuber, which described linear transformers using the delta rule as associative memory, or “fast weights.” He presents Qwenputer’s contribution as making such memory exact and addressable in a production-model architecture, then compiling a CPU to use it. The blog says no training is involved: a compiler sets the added weights, while the original Qwen weights remain unchanged.

““The CPU runs completely within the forward passes of the model.””

— Divyansh Agrawal, Qwenputer project report

Independent Testing and Limits

The performance figures and compatibility claims come from Agrawal’s project report. The source material does not provide independent replication, detailed benchmark conditions, or a separate audit of the implementation. It also does not quantify execution speed, memory capacity, or how performance changes across different programs.

The report says the system can run C and MicroPython “quite well” but also describes it as slow; it gives no timing data. It is not clear from the supplied material how broadly the approach works beyond this specific Qwen3.5 4B build, or what practical limits arise when programs grow. The description of the CPU as deterministic is a project characterization; the report’s test results support repeatability claims, but do not establish every possible behavior under every workload.

Code and Model Availability

Agrawal says Qwenputer is available through Hugging Face and ModelScope, with code on GitHub. Interested researchers can examine the implementation and attempt to reproduce the reported tests. The source material does not identify a scheduled release, independent benchmark, or formal evaluation as the next milestone.

The clearest next step for assessing the project is independent testing: verifying the RISC-V results, measuring execution speed and memory limits, and checking whether the design can run a wider range of programs. Until those details are available, Qwenputer is best understood as a reported proof of concept showing a CPU compiled into a modified language model—not as a general-purpose processor or an inherent capability of stock Qwen3.5.

Key Questions

What is Qwenputer?

Qwenputer is an experimental modified Qwen3.5 4B model with a RISC-V CPU compiled into additional weights. Its CPU runs through model forward passes rather than through an external execution tool.

How does it store CPU memory?

Agrawal uses Gated DeltaNet state matrices as addressable storage. In the described setup, one-hot keys select rows for writes and reads, allowing the CPU to store values such as registers and its program counter.

What does the CPU reportedly support?

The report describes an RV32IM core with the Zba extension and a small bignum unit. Agrawal says it passes official RISC-V tests, matches QEMU on programs, and can run C and MicroPython, though slowly.

Has the 18-million-instruction result been independently verified?

Not in the supplied source material. Agrawal reports that a free-running test completed 18 million instructions without error, but no independent replication or external audit is described.

Is this a feature of stock Qwen3.5?

No. Qwenputer is a modified experimental build with 11 added computer layers. Agrawal says the original Qwen weights remain unchanged beneath those additions.

Source: divagr.com

You May Also Like

Acoustic Dampening, Placement, and the “Rig in the Closet” Setup

Learn how to reduce noise from high-power AI workstations through placement, acoustic treatment, and the ‘rig in the closet’ method, with expert insights.

Explore 2026’S Best AI Camera Lenses For Every Photographer

Explore the best AI camera lenses for 2026, tailored for different photography needs, with expert insights on features, compatibility, and value.

US Government Bans An Open Source AI Model In 2026?

The US government has officially banned an open source AI model in 2026, citing national security concerns. Details remain limited, and the decision’s implications are still unfolding.

Bitcoin Battles Unfold in Live Warzone Visualization

A new browser-based visualization depicts the live BTC/USDT trading battle as a cinematic war scene, highlighting market dynamics without trading advice.