The experimental Qwenputer system configures a stock Qwen 3.5 4B model to house a fully functional 32-bit RISC-V processor that executes entirely within token generation steps. Rather than relying on traditional machine learning training, engineers utilized a compiler to set network weights directly from circuit descriptions across 11 layers. To handle exact computational state, the setup repurposes the architecture's linear RNN layers into addressable random-access memory and experimental file storage. The resulting virtual processor successfully runs official CPU test suites and produces step-by-step execution traces matching standard emulators.
Key points
Neural network weights can be compiled directly from circuit descriptions to run traditio…: Neural network weights can be compiled directly from circuit descriptions to run traditional code deterministically without machine learning training.
Utilizing linear RNN states as addressable memory provides a lower computational cost tha…: Utilizing linear RNN states as addressable memory provides a lower computational cost than relying on standard key-value caches.
Embedding computation within hidden recurrent states raises safety concerns regarding byp…: Embedding computation within hidden recurrent states raises safety concerns regarding bypassed chain-of-thought monitoring alongside opportunities for untamperable auditing.
Sources checked by our analysis pipeline: 1
COMPILED WEIGHTS
Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5
A stock Qwen 3.5 4B with a RISC-V CPU compiled into its weights, running on its own Gated DeltaNet layers as RAM.
I can't believe it's not silicon
Qwenputer is a pretty fun experimental model. It's a stock Qwen 3.5 4B with one addition: a RISC-V CPU embedded within the weights that uses the model's own Gated DeltaNet attention layers as RAM. (To learn how GDN works, here's a handy guide.) This 'CPU' runs completely within the forward passes of the model, and can run arithmetic operations, C and MicroPython code without any external runtime or tool usage whatsoever.
To be very clear, it is not a plugin, a tool or a sandbox. This blog will explain exactly how this model works, and how the model utilises GDN layers as RAM, and how I also slotted in an experimental in-model file storage.
I also wrote up two follow-ups built on top of Qwenputer: how hybrid models can learn to hide their reasoning in their RNN state and evade CoT…
The technology giant is offering its most intelligent frontier model to specialists aiming to strengthen their digital security postures. Early access provides a temporary cost advantage for organizations testing the system before service fees double. This specialized deployment targets critical defense capabilities using tailored artificial intelligence infrastructure.
Advanced technology from leading American developers continues to find its way into competing overseas development pipelines. Regional artificial intelligence organizations in China have integrated OpenAI's Astra into their workflows to build and train their proprietary systems. This activity highlights the ongoing reliance on foundational Western architectures to enhance foreign machine learning capabilities.
Investor interest in automated software development continues to escalate rapidly as venture capital firms compete for stakes in category leaders. The software development startup is evaluating fresh capital proposals that represent a dramatic surge over its prior benchmarks. Securing a deal at these proposed levels would position the firm among the most valuable private software ventures globally.